# LADLE — full corpus > Frontier intelligence. Powered by Claude. LADLE is a general-purpose AI assistant powered by Claude Sonnet, with web search, files, images, and code — same class of model you already pay for. Of every $20 subscription, $8 buys meals through the UN World Food Programme at $0.80 each. Receipts published monthly with reference numbers. If we ever can't show the receipt, cancel us. ## About LADLE PBC is a Delaware public benefit corporation building a Claude-powered AI assistant subscription. Every subscription commits at least 10 meals per cycle to the UN World Food Programme through ShareTheMeal at $0.80 per meal. ## Plans **Ladle** — $20/mo, 10 meals/mo via WFP. The one plan most people should pick. **Ladle Max 5x** — $100/mo, 50 meals/mo via WFP. Heavier use, longer files, priority at peak. **Ladle Max 20x** — $200/mo, 100 meals/mo via WFP. Rate limits become a thing you don't think about. ## Blog posts (56) ## Six months in. - Slug: six-months-in-honest-retrospective - URL: https://ladle.chat/blog/six-months-in-honest-retrospective - Date: 2026-08-08 - Category: NOTE - Summary: March to August. What worked, what didn't, and the two things I got embarrassingly wrong. Not a growth post — a lessons post. Six months since the March launch. This is the retrospective. Not a growth-numbers post — we'll write those at year-end if we're still growing enough for the numbers to be interesting. This is a lessons post. **What worked:** **The receipts.** The single strongest signal that the whole approach works has been the reception to the monthly WFP receipt. Users share screenshots of them, prospects mention them as the reason they subscribed, journalists writing about consumer AI have referenced them. The receipt turned out to be the marketing, exactly as we'd hoped. **Not chasing free.** Every month of not launching a free tier has been a month of not having to defend the meal donation math against "but why don't the free users get meals too." The demo (three messages on marketing, no signup) has been enough. Zero regret on this call. **Being specific in copy.** The marketing copy says "$8 buys ten meals through WFP's ShareTheMeal partner at $0.80 each." Not "we donate a portion." Not "we give back." Specific numbers, specific partner, specific per-unit cost. This has been our biggest differentiator against generic-impact marketing from every other consumer subscription trying to look ethical. **Kitchen panel and model transparency.** We shipped this without knowing if users would care. They do — the most engaged subscribers use it, mention it to friends, and treat it as one of the reasons LADLE feels different from the black-box competitors. **What didn't work:** **The first onboarding.** The three-step "get to value fast" version we launched with produced worse retention than the ten-step "Setting the Table" we shipped later. Should have started with the longer version and let it prove itself, rather than shipping the industry-default shortest one first and rebuilding. **Design mode's first version.** Before the canvas-first rebuild in August, design mode was just chat with a different system prompt. Nobody used it meaningfully. The lesson: system-prompt-only features don't feel like features to users — they need distinct surfaces. Should have done the canvas from the start. **The billing hesitation.** We spent longer than we should have without shipping Stripe. The market for "consumer AI subscription with meals attached" was ready in June; we didn't have real checkout until August. Two months of coasting on a placeholder page is two months of lost momentum. Lesson: ship billing early even when it's ugly, then polish. **Two things I got embarrassingly wrong:** **One: I assumed users would want fewer settings, not more.** The original composer had four toggles hidden behind a menu. Users kept asking where the toggles were. Turns out for a $20/month product, users want visible controls, not "we've hidden the complexity for you." Every toggle got promoted to the composer surface after month two. Should have started that way. **Two: I underestimated how much the community would care about the specific WFP partnership.** I picked WFP for the operational reasons (scale, receipt clarity, per-meal math). What surprised me was how often subscribers wrote in with a personal story about a family connection to food insecurity, or a professional connection to humanitarian work. This wasn't in the market research. It should have been. **What I'm watching for the next six months:** Whether the receipts model scales past a certain subscriber count without losing its intimacy. At 500 subscribers, every one is close-to-personal; at 50,000 the receipts become anonymous line items. The lesson from month one publishing carefully has to hold at 100× scale. Whether the "no free tier" position gets tested by cheap-model competitors. DeepSeek's essentially-free tier is real competition on price. So far the value prop has held, but the market is dynamic. Whether we can keep the changelog cadence tight enough to matter. Silent weeks would be visible in a way most companies don't have to worry about, because we've committed to the changelog being the marketing. **One thing I want to say explicitly:** The subscribers who joined in the first three months when the product was rougher — thank you. That's the group that took the promise at face value before there was much of a track record. Every month since has been building the track record. The receipts page is the ongoing thank-you. If you're reading this having just subscribed: hello, thanks, and if something feels wrong please write. If you're reading this after canceling: also thanks, and honestly, no hard feelings — the refund policy exists for exactly the "not for me" case. Six months. Onward. --- ## LADLE vs 'just use the Anthropic API directly.' - Slug: ladle-vs-just-use-the-api - URL: https://ladle.chat/blog/ladle-vs-just-use-the-api - Date: 2026-08-07 - Category: NOTE - Summary: Yes, you could hit Claude's API yourself and pay pennies per turn. Whether that math beats a $20/mo product depends on what you actually value. The most technically-honest question anyone asks about LADLE: "Why pay $20/mo when I could just call the Anthropic API directly for way less?" Let's do the math. **The API cost side.** Claude Sonnet on Anthropic's API is $3/MTok input, $15/MTok output (verified 2026-08-09). A typical chat turn — 500 input tokens + 200 output tokens — costs roughly $0.0015 + $0.003 = $0.0045. At 300 turns/month (heavy daily use), you'd spend $1.35 raw. Adding prompt caching brings it down further. **The API friction side.** You'd need: an Anthropic account with billing, API credentials in your environment, a client library, a way to persist chats (SQLite? Postgres?), a UI (build your own? Cursor? Claude Desktop?), memory/instructions handling, projects, artifact storage, and a way to attach files. All buildable; all real work. **The gap between "API access" and "finished product."** The API gives you the raw model. Everything else — the UI, the storage, the file handling, the design canvas, the kitchen panel, the shared surfaces, the mobile experience, the auth, the payment infra — you build yourself or forgo. **When the API path is right.** - You're a developer with time to build (or an existing chat product to plug Claude into). - Your usage is very low (a handful of turns per week) — a $20/mo subscription is overkill. - You need programmatic access from your own scripts/tools (LADLE is a chat product, not an API). - You want to build something specifically not-available in any consumer product. **When LADLE (or Claude Pro, or ChatGPT) is right.** - Your time is worth more than the $18-19/mo delta above raw API costs. - You value a finished product over building it. - You use AI conversationally more than programmatically. - You want features that require infrastructure (Projects, Memory, chat history, mobile). **The LADLE-specific answer.** If you're going to buy a finished chat product anyway, LADLE's $20 sends $8 to WFP with a monthly receipt. If you're not going to buy a finished chat product — you're going the API route — LADLE isn't the right recommendation and I'd rather you use the API and donate to WFP directly than pay us reluctantly for infra you don't need. **Honest self-assessment.** If reading this list makes you realize "I could probably build my own chat interface in a weekend," you probably should. LADLE isn't the right product for you. Come back when the build/maintain math flips. --- ## LADLE vs Claude Pro — the exact overlap. - Slug: ladle-vs-claude-pro-the-overlap - URL: https://ladle.chat/blog/ladle-vs-claude-pro-the-overlap - Date: 2026-08-06 - Category: NOTE - Summary: Two products on the same model at the same price. Here's the feature-by-feature reality — where they overlap identically, where LADLE lags Claude Pro, and where LADLE has its own opinions. The most direct competitor comparison LADLE has. Anthropic's own Claude Pro is $20/mo, runs on Claude Sonnet, and offers substantially the same core capabilities. Here's the feature-by-feature honest read. **Identical or near-identical.** Underlying model (Claude Sonnet 4.6). Model behavior on all core tasks. 200K context window. File uploads and image inputs. Web search when enabled. Extended thinking. Basic chat surface + Projects. Pricing. **Where Claude Pro has more (as of 2026-08).** - Native integrations with Google Drive, Notion (LADLE hasn't wired these). - Anthropic ships new features to Claude.ai first; LADLE integrates on a lag of days-to-weeks. - Computer use (agentic model that operates a browser/GUI). LADLE hasn't shipped this. - Direct access to Anthropic Console (build with the API alongside chat use). **Where LADLE has more.** - The meals. $8 of every subscription funds meals through the WFP, published receipt monthly. Claude Pro's $20 stays entirely in Anthropic's economy. - Kitchen panel (transparent routing/token per turn, Cmd+K). Claude.ai has less of this visible. - Design canvas (chat + iframe preview + inspector-driven scoped edits for HTML/design work). - Explicit response-style settings (Concise / Thorough / Exploratory / Precise) as first-class toggle. **The wash.** Chat UX quality is subjective; both are considered products. Speed is roughly equivalent. Reliability is equivalent (same underlying model + API). **When to pick Claude Pro.** You want the newest Anthropic features the day they ship. You want the Google Drive / Notion integrations. You're already using the Anthropic Console for API work and want one account. You don't care about the meal donation. **When to pick LADLE.** You want the meal donation as part of your $20/mo. You value the specific product opinions (kitchen panel, design canvas, response-style toggle, honest error copy). You want a chat product that's opinionated about transparency. **Neither is objectively better.** They're the same class of tool with different priorities. Try both for a week if you're deciding; the model behavior is identical so the choice is really about product surface and social contract. --- ## What comes next — and what doesn't. - Slug: what-comes-next-and-doesnt - URL: https://ladle.chat/blog/what-comes-next-and-doesnt - Date: 2026-08-06 - Category: NOTE - Summary: The next-few-months build list, the not-building list, and the specific things that would move an item from the second list to the first. This post is a public roadmap. Not the "here's our vision for 2028" version — the "here's what we're actually working on for the next few months" version. And, importantly, the list of what we're deliberately NOT building. **Coming in the next few months (in rough order):** **Mobile apps.** LADLE works well in mobile browsers today, but a native iOS and Android app would improve the daily-use case for a lot of subscribers. iOS first. No timeline commitment; when it ships, the changelog gets a line. **Better projects.** Projects work today (chats grouped, shared instructions, knowledge files at the project level) but the UI is under-considered. A rework is queued. **Team seats with light admin.** Not a full "team product" — just: buy multiple seats on one card, one billing contact, one impact report. No shared workspace. This unblocks the small-business use case without committing to the full team-product roadmap. **Referral / affiliate.** The /creators page describes a program that doesn't exist yet. Building it — flat rate per referral, honest math, publicly documented. **Live model updates.** When Anthropic ships a new Sonnet or Haiku version, we need a smoother update path than "we redeploy and users find out from the changelog." Model updates deserve their own release notes. **Impact API.** Some business customers have asked if they can pull their organization's monthly impact numbers via API for their own dashboards. Low-effort to build; queued. **Not building (for now):** **Image generation.** Diffusion models are outside our stack and outside our value prop. If you need image gen, dedicated tools do it better. **Voice mode.** The audio chat pattern is nice for some users. It's a lot of surface area to build well — audio streaming, voice selection, mobile audio state management — and we don't currently see it as our core. Might change if enough subscribers ask. **Custom GPTs / agent builder.** OpenAI has GPTs; Anthropic has Projects (which we support). A custom-agent-builder inside LADLE is a big product that would distract from the core. Not on the list. **Enterprise SSO / SCIM.** No demand from subscribers yet. If a big-enough deal shows up requiring it, we build it for that deal. **A public API for LADLE itself** (letting developers hit LADLE's chat endpoint from their own apps). Interesting but not core; Anthropic's API is the answer for developers who want to build with Claude directly. **A LADLE-branded model, fine-tuned or otherwise.** We use Claude Sonnet as-is. No LADLE-specific training data, no proprietary fine-tune. The model is Anthropic's; the product is ours. This is deliberate. **What would move something from the "not building" list to the "building" list:** Repeated specific asks. Not "would be cool if" — actual "we would sign up / stay subscribed / recommend if" requests. When we see the same request from a few different people who match the customer we're building for, that's signal. A specific unlocked use case. If enterprise SSO would unblock a 200-seat customer, we build it for that customer. A material shift in the underlying tech. If open-source models close the gap to Sonnet at meaningfully lower cost, the economics could support things that don't fit today. **How we prioritize:** The order isn't "biggest feature first" or "loudest request first." It's roughly: (1) things that fix specific broken user experiences, (2) things that unblock specific stuck subscribers, (3) things that expand the reachable-users set materially, (4) new capabilities. Meal-donation infrastructure doesn't compete with other work; it always ships first when it needs attention. That's baked into how we run. **What NOT to expect:** Big splashy quarterly launches. We ship small and often. If we shipped a big feature quietly, it made it into the changelog with a line — the changelog is the announcement. A roadmap document that's revised quarterly with dates. Our roadmap is the changelog looking backward and this post looking forward. The dates are "when it ships." Pre-announcements of features six months out. If we're not going to ship it in the next quarter, we're not going to pre-announce it. **Rule of thumb.** If you want to know what LADLE is doing, read the changelog for the last two months and this post for the next two. That's the whole picture, and the picture updates as we go. --- ## Design mode was a two-day pass. - Slug: design-mode-two-day-pass - URL: https://ladle.chat/blog/design-mode-two-day-pass - Date: 2026-08-05 - Category: NOTE - Summary: LADLE Design turned the chat surface into a canvas — iframe preview left, chat rail right, click-to-select for scoped edits. Here's what shipped and what got left for later. We shipped LADLE Design as a two-day pass in early August. The scope was deliberately narrow: make the design workflow feel like a design tool, not a chatbot with a preview panel. Before this pass, hitting "New design" on LADLE just opened a chat with a different system prompt. Same UI as everything else. The reply, if the model decided to produce HTML, would land in an artifact panel you could open on the side. Technically it worked. In practice, "design" didn't feel like design — it felt like chat with a hidden output. **What shipped in the split.** The canvas surface at /design/[chatId] flips the ratio. The iframe preview takes 65% of the screen. The chat rail takes 35% and lives on the right. Every reply auto-updates the iframe. Below the preview: a strip of thumbnails for every previous version in the chat, one-click to revert. Above: device chips (desktop / tablet / phone) so you can see how the design looks at different widths without leaving the surface. The "New design" CTA on /design/new opens the empty canvas with the composer focused. Type a brief, hit send, watch the design render live. **What shipped in the inspector.** The Select toggle in the canvas header injects an inspector script into the sandboxed iframe. Turn it on, click any element in the preview, and its CSS selector jumps into a chip above the composer. Type a change — "make this button bigger," "swap this copy" — and the request is scoped: the server augments the prompt with a SELECTION header, routes to Haiku for a sub-second edit, and returns a new version with everything else preserved. Scoped edits go to Haiku because they don't need Sonnet's depth — they need surgical precision. Full designs still route to Sonnet. The user doesn't have to choose; the presence or absence of a selection is the signal. **What shipped in Publish.** A Publish button snapshots the current version to a public URL at /d/[slug]. Copy the URL, paste it anywhere, and the design renders fullscreen with a small "designed with ladle" credit in the corner. Auth-free for readers. The snapshot is immutable — editing the artifact after publishing doesn't change what the public URL shows. **What didn't ship.** Undo/branch pinning. Right now versions are linear — a new reply appends. Pinning a version as a branch base for iteration is on the list. Inline diff pulse. When an element edit lands, the changed element should briefly glow to confirm what moved. Didn't get to it. Kitchen budget line inside the canvas. Cost estimate per turn would help heavy users manage their cycle. Data is captured; UI is deferred. **The point of the split.** Before: LADLE Design looked like ChatGPT with a maybe-preview. After: it looks like v0 or Claude Artifacts — a working canvas with chat as the crew, not the star. That reframing is worth more than any specific feature we added around it. If you tried Design before this pass and bounced, try it again — it's a different product now. Blog is the receipt. Design shipped 2026-08-05 to 2026-08-07 across two sessions. The changelog has the day-by-day. --- ## LADLE vs DeepSeek — if you care where the money goes. - Slug: ladle-vs-deepseek-money-goes - URL: https://ladle.chat/blog/ladle-vs-deepseek-money-goes - Date: 2026-08-04 - Category: NOTE - Summary: DeepSeek is essentially free and genuinely competitive on model quality. The comparison isn't features — it's whether the social contract of your $20 matters to you. DeepSeek changed the AI subscription landscape in a way that deserves an honest response. The chat product is free at the tier most people use, the model is genuinely competitive on many tasks, and the value proposition to a cost-sensitive user is significant. Here's the honest LADLE-vs-DeepSeek framing. **When DeepSeek is the right choice.** You want capable AI for everyday tasks and don't have strong preferences about who runs the model or where the money goes. DeepSeek's free tier covers most usage patterns. Your $20/mo saved from LADLE can be spent on other things — including, if you want, donating directly to WFP through ShareTheMeal for the equivalent of what LADLE would have given. **When LADLE is the right choice.** You care where the money goes. LADLE sends $8/mo to WFP with a public receipt. DeepSeek's free tier sends $0 anywhere — value flows into DeepSeek's business, which is legitimate but doesn't do the specific meal-donation thing you'd have to do yourself. **Model provenance matters for some users.** DeepSeek is a Chinese-based company. If your work touches sensitive material or you have regulatory constraints (many US and EU enterprises do), the operating jurisdiction is a real consideration. LADLE runs on Anthropic (US-based). Neither situation is inherently better; the answer depends on your specific data-handling requirements. **Model behavior differences.** DeepSeek's model is competitive but not identical to Claude Sonnet — different training, different alignment defaults, different response style. If you've settled into Claude's way of writing and thinking, switching to DeepSeek is a real behavioral adjustment. **The honest positioning.** LADLE is not competing with DeepSeek on price. We can't and don't want to. We're competing on what happens with the money you spend on AI, and on the specific product experience we ship. If price is your primary axis, DeepSeek wins. That's a legitimate choice. I'd rather you use DeepSeek and donate the $20/mo you save to WFP directly than reluctantly pay LADLE and resent it. If price is NOT your primary axis — you're spending $20/mo on AI anyway, and given the choice, you'd like the software you use to have a specific and unusual social contract — then LADLE. That's who we're for. --- ## Should you use LADLE or DeepSeek? - Slug: ladle-or-deepseek - URL: https://ladle.chat/blog/ladle-or-deepseek - Date: 2026-08-03 - Category: NOTE - Summary: DeepSeek is functionally free. LADLE is $20/month with $8 going to the WFP. Both are legitimate choices. The right one depends on why you're paying (or not) in the first place. DeepSeek's release changed the landscape. The chat product is essentially free — you can use it heavily without hitting a paywall. The model is genuinely competitive on many tasks. Some people have asked us how LADLE's $20/month justifies itself against a free alternative that works. Fair question. Here's the honest answer. **When DeepSeek is the right choice:** You want cheap AI for everyday tasks and don't have strong preferences about who runs the model or where the money goes. DeepSeek is free at the tier most people use, competent for most everyday work, and the trade-offs (see below) don't matter to you. You're cost-sensitive on principle — $20/month across 10 subscriptions adds up, and you'd rather aggregate on the free option and save the $240/year for other things (including, if you want, donating directly to WFP through ShareTheMeal for the equivalent of what LADLE would have given). **When LADLE is the right choice over DeepSeek:** You care where the money goes. LADLE's $20/month sends $8 to WFP monthly with a public receipt. DeepSeek's free tier sends $0 anywhere — the value goes back into DeepSeek's business, which is fine, but doesn't do the WFP thing you'd have to do separately. You care about model provenance. DeepSeek is a Chinese-based company. If your work touches sensitive material (business strategy, personal information, legal documents), or if you have regulatory constraints (many US and EU companies do), where the model is operated matters. Anthropic is US-based; their data policies are well-known. Do your own due diligence, but the provenance question is a live one for some users. You want the specific model behavior of Claude Sonnet. DeepSeek's model is competitive but different — different training, different alignment approach, different response style. If you've settled into Claude's way of writing and thinking, switching to DeepSeek is a real adjustment. You value the product opinions LADLE ships (kitchen panel, design mode, artifact-first workflow, transparent routing). DeepSeek's chat interface is functional but bare. **When neither is the right choice:** You need enterprise-grade features (SSO, audit logs, dedicated support). Neither LADLE nor DeepSeek is the fit here. Anthropic Enterprise, Claude Team, or OpenAI Enterprise are the categories. You need cutting-edge frontier model capability for hard tasks. Claude Opus, GPT-5, or Gemini Ultra at the top tiers are the answer. LADLE (Sonnet) is one step down from the frontier tier for cost reasons; DeepSeek's flagship is close to frontier but with the provenance caveat. You need very specific integrations (Google Drive, Notion, custom internal tools). Anthropic direct or OpenAI have more integration surface than LADLE currently does. **The honest positioning:** LADLE is not competing with DeepSeek on price. We can't; we don't want to. We're competing on what happens with the money you spend on AI, and on the specific product experience we ship. If price is your primary axis, DeepSeek wins. That's a legitimate purchase decision. We'd rather you use DeepSeek and donate the $20/month you save to WFP directly than reluctantly pay LADLE and resent it. If price is NOT your primary axis — if you're spending $20/month on AI anyway, and given the choice, you'd like the software you use to have a specific and unusual social contract — then LADLE. That's who we're for. **Rule of thumb.** The choice isn't LADLE vs. DeepSeek. It's "what are you optimizing for." Cheapest? DeepSeek. Best frontier reasoning? Different tier entirely. AI subscription that does something specific with the money? LADLE. Nothing about this needs to be adversarial. There's room for multiple legitimate answers to "which AI should I use," and the right one for you depends on what you're actually solving. --- ## LADLE vs Gemini — for document-heavy work. - Slug: ladle-vs-gemini-for-docs - URL: https://ladle.chat/blog/ladle-vs-gemini-for-docs - Date: 2026-08-02 - Category: NOTE - Summary: Gemini's Google Workspace integration is real advantage for document work if you live in Docs/Sheets/Slides. LADLE handles docs well via upload but doesn't integrate with them live. If your document work happens in Google Workspace (Docs, Sheets, Slides, Gmail), Gemini has a structural advantage LADLE doesn't try to match: live integration inside the tools you already use. Here's the honest tradeoff. **Gemini's document integration.** Gemini for Workspace ($20-30/user/mo depending on tier) sits inside Docs, Sheets, Slides, and Gmail. You can ask it to draft a section of the current doc, summarize a spreadsheet you're viewing, or draft a reply to the email you have open — with the current document as context, no copy-paste. For heads-down Workspace users, this is a real time-save. **LADLE's document handling.** You upload the file (drag into composer or attach), ask questions in chat, get responses. Same core capability without the integration — you're moving content between LADLE and your document rather than working in one surface. For a summary or an edit, this is fine. For iterative drafting inside the doc, it's more friction. **Where LADLE wins on docs.** Long-context recall on very large documents (200K tokens is roughly 500 pages). Gemini has strong long-context handling too, but LADLE's chat surface + Projects structure is often cleaner for research-across-many-docs workflows. Also: LADLE's editing quality on prose tends toward stronger literary sensibility, useful for polished writing. **When Gemini fits better.** You live in Google Workspace. Your work is repeatedly "draft this section," "summarize this sheet," "reply to this email" — inside those tools. Gemini's cost adds meaningful integration value for that specific pattern. **When LADLE fits better.** You work across multiple document tools (Notion, Slack, PDFs, whatever), or you export to a document tool but do the thinking in a chat surface. Also: if you specifically want the $8/mo meal donation as part of your $20 subscription. **Neither is strictly better.** The document-integration story favors Gemini for Workspace-native workflows; the standalone-chat-with-strong-writing story favors LADLE for polyglot workflows. Try Gemini's free tier alongside a LADLE trial for a week and see which surface you actually reach for. --- ## LADLE vs ChatGPT — for daily driver use. - Slug: ladle-vs-chatgpt-for-daily-use - URL: https://ladle.chat/blog/ladle-vs-chatgpt-for-daily-use - Date: 2026-08-01 - Category: NOTE - Summary: The daily-driver comparison. Both handle 95% of everyday AI use equally well. Choose on chat surface preference, model voice, and the social contract of your $20/mo. The most-asked comparison. For "which AI subscription should be my daily driver," here's the honest answer. **Where they're basically identical.** Model quality on everyday tasks (drafting emails, summarizing documents, answering factual questions, code snippets, brainstorming, meeting notes). Both handle files. Both have Projects. Both have image input. Both have web search. Both are $20/mo for the base tier. **Where they differ meaningfully:** **Model voice.** Claude (LADLE's model) tends toward more thorough, occasionally more literary replies. GPT tends toward more matter-of-fact, tighter defaults. Neither is better; they're different flavors. If you loved GPT's clipped defaults, Claude will initially feel chatty (fix: LADLE's "Concise" response style). If you loved Claude's care, GPT will feel abrupt. **Chat surface.** LADLE has a kitchen panel showing which model produced each reply and why (Cmd+K). ChatGPT hides its routing. LADLE has a design canvas (chat + iframe preview + inspector) for HTML/design work. ChatGPT has better image generation (DALL-E integrated). Preference-driven. **Extras.** ChatGPT has voice mode (spoken conversations), advanced data analysis (Python in a sandbox), better custom-GPT ecosystem. LADLE doesn't have those. LADLE has the artifact-first workflow, more transparent routing, and explicit design mode. **What's structurally different.** The reason LADLE exists as a separate product from Claude.ai (Anthropic's own $20/mo direct offering) is that $8 of every LADLE subscription funds meals through the World Food Programme, with a monthly public receipt. ChatGPT's $20 stays entirely in OpenAI's economy. **Recommendation for daily use.** - If you're already deep in ChatGPT and it works, stay put. Switching costs (custom instructions, projects, workflow habits) are real; the model quality delta is small. - If you're new, try both for a week each and pick on chat-surface preference. - If you care about the meal donation and you're at parity on features, LADLE. - If you specifically need image generation or voice mode, ChatGPT — LADLE doesn't have either yet. --- ## Who Max 20x is actually for. - Slug: when-max-twenty-makes-sense - URL: https://ladle.chat/blog/when-max-twenty-makes-sense - Date: 2026-08-01 - Category: NOTE - Summary: Max 20x isn't Max 5x with a bigger number. It's a different product for a specific kind of user — one whose day materially depends on the model. Here's how to tell. Max 20x is our largest plan: $200/month, 100 meals per cycle, effectively unlimited practical usage for a single power user. Most people should not buy it. This post is about the specific people who should. Start with usage. Ladle base gives you generous daily use — hundreds of chats a month, most files, most workflows. Max 5x is five times that. Max 20x is twenty times. On paper, "20x" means "hits rate limits basically never." In practice, the profile of someone who USES 20x that much is narrow. Three profiles we've observed in the first months: **The writer whose draft loop lives in the model.** Not "I use AI to help me write." I mean: your first pass gets a full read from the model, you rewrite based on the note, you paste it back for another read, and this loop happens 12-25 times per essay. If that's your Tuesday, Max 5x might catch you at peak and Max 20x won't. **The researcher running document pipelines.** You have 200-page PDFs, RSS feeds, court filings, corporate reports. You extract, summarize, cross-reference, and repeat. This is not "chatting with docs" — it's using the model as the middle step in a workflow that generates its own next input. Max 20x here reflects "this is how I earn money." **The engineer using Claude Code adjacent workflows.** Long context reads of full repos, exploratory refactors, generated tests, iterative debugging — all the workflows that individually use a lot of tokens. Not "occasionally paste code." I mean: 4-8 hours a day, the model is the second developer. If none of those three sound like you, Max 5x is almost certainly enough. Some tell-tales for staying on Max 5x or Ladle base: - You use AI for "a few things a week" or "as needed" → Ladle base is the answer. - Your peak usage is one intense day per week → Max 5x, and you probably won't hit its ceiling either. - You want the meal count to be higher and don't need the capacity → the honest answer is: don't buy capacity you won't use. Buy Ladle base and top up on gift subscriptions instead. Same meal impact, no waste. The meal math on Max 20x: 100 meals per subscription per cycle, $80 to WFP. That's a real donation. But it's the same $0.80/meal math as every plan — Max 20x doesn't multiply the per-meal impact, it multiplies the count because you're paying more. The trap on high tiers is buying capacity for the aspirational version of yourself. If Max 20x sits idle, you didn't help WFP more than you would have on Ladle base for the same capacity you actually used. You just spent more. If you're on Max 20x and looking at your usage graph in Settings, and it's below 40% of Max 5x's capacity for three cycles running, downgrade. We'll never nudge you to upgrade beyond your usage; we'd rather have you long-term at the tier that fits. The right tier is the one you actually use. --- ## Why we publish receipts. - Slug: why-we-publish-receipts - URL: https://ladle.chat/blog/why-we-publish-receipts - Date: 2026-08-01 - Category: NOTE - Summary: The one operating decision that shapes everything else about how LADLE works — and the reason we have no ad budget, no fundraising deck, and no retention team. The most important thing on this website is the ledger page. Not the landing page, not the demo, not the comparison tables. The ledger — dates, amounts, meal counts, reference numbers from ShareTheMeal, published on the day each transfer clears. Everything else on the site is downstream of that decision. We could operate LADLE without publishing anything. Pick a nice-sounding fraction ("a portion of every subscription goes to good causes"), route the money quietly, and never have to prove it. The competitive set does this, mostly. A public commitment that isn't tied to a receipt is really just a mood. We didn't want to run that company. The whole point of LADLE is that you're already paying $20 a month for a general assistant, and if some of that money is going to do a second job on the way out, you should be able to verify it. Not eventually, not with a login, not with a "sustainability report" published quarterly at PDF page 34 — verify it, on a URL, with a reference number you can email ShareTheMeal about. That commitment has three second-order effects that shape the rest of the business. First: it caps how creative we can get with the price. The moment we start experimenting with a Pro tier at $30 or a Team tier at $50/seat, the meal math gets more complicated to publish honestly. So we don't. One product, one price. Second: it kills the retention team playbook. Every SaaS company has some version of "cancel? Are you sure? Here's 50% off." Ours can't. If the pitch was "you're paying for meals to be funded", offering a discount on cancel means saying "you're funding fewer meals now, but still paying us" — which is worse than just letting you go. So the cancel screen is one button. Third: it makes marketing straightforward. We don't have to invent a story about impact. We just link to the ledger. The receipts do the work that a testimonial section would do at another company. This is also why we don't have a Series A pitch deck. What we have is a spreadsheet: subscribers × $8, minus API cost, minus ops, minus $0 for growth marketing because we don't run any. It either works at scale or it doesn't. If it does, we grow. If it doesn't, we don't get to reinvent ourselves into a different company that's better at fundraising. The receipts, and the fact that the receipts are the whole trust mechanism, are the operating decision. Everything else is a consequence. --- ## LADLE vs Copilot — for coding work. - Slug: ladle-vs-copilot-for-code - URL: https://ladle.chat/blog/ladle-vs-copilot-for-code - Date: 2026-07-31 - Category: NOTE - Summary: Copilot completes code in your editor. LADLE is chat-first — code review, refactoring, debugging. They coexist better than they compete. A common confusion: developers ask whether to use GitHub Copilot or LADLE for coding. The honest answer is that they're solving different problems. **Copilot** ($10-19/mo per user) lives in your editor and completes code as you type. Its strength is inline suggestions — you start typing a function signature and it autocompletes a plausible implementation. It's a productivity multiplier on the mechanical work of typing. **LADLE** ($20/mo) is chat-first. You paste code, ask a question, discuss the reply, iterate. The output is prose + code together. It's a productivity multiplier on the thinking-about-code work — reviewing a diff, deciding between refactor approaches, debugging a subtle issue, explaining a codebase you're new to. **Overlap that matters.** Both can generate code from a description. Copilot does it faster (inline) but shallower. LADLE does it slower (chat) but with far more context (paste the whole file, explain the constraints, iterate). For a small function, use Copilot. For a component, module, or refactor: LADLE. **When neither.** For the parts of coding that require deep tacit knowledge of a large codebase, neither tool fully substitutes for a senior engineer who's read the whole thing. Both help. Neither replaces. **Common pairing.** Use both. Copilot for typing velocity, LADLE for thinking. Combined cost: ~$30-40/mo. If your work is mostly heads-down implementation of well-specified tickets, Copilot alone is fine. If your work is mostly design/review/debugging with implementation as a smaller piece, LADLE alone is fine. If both, both. **What's structurally different.** LADLE's $20 sends $8 to WFP. Copilot's fee stays in GitHub/Microsoft's economy. If the difference matters, LADLE. If not, either. **One thing to know about Copilot for code.** Copilot Business/Enterprise tiers include additional data-handling guarantees (your code isn't used for training). Check your organization's policy before enabling personal Copilot on work code. --- ## Support at a team of one. - Slug: support-at-a-team-of-one - URL: https://ladle.chat/blog/support-at-a-team-of-one - Date: 2026-07-28 - Category: NOTE - Summary: Every email to hello@ladle.chat lands with the founder. Here's how that works today, what it costs in hours, and the specific triggers that would change it. LADLE is a team of one. Product, code, marketing, ops, and support all funnel to the same person. This post is specifically about support — how it's structured today, what it actually costs in hours, and what would change it. **How it works today:** Every email to hello@ladle.chat lands in the founder's inbox. Every one is read and answered by the founder personally, usually within 24 hours, often within a few. There's no ticketing system, no support tier, no first-line agent. The email you send is the email that gets read. Same for the other support routes: security@ (rare, thankfully), partners@, press@, privacy@. All routes end at the same person. **Why it's structured this way:** At current scale, the volume is manageable — a handful of support emails per day, most of which are billing questions ("how do I cancel," "when does my renewal hit") or product questions ("how does the search toggle work"). A handful of technical bugs per week. Occasional larger conversations. The founder-in-the-inbox model has two advantages worth defending as long as possible: **Speed of feedback loop.** When a bug report lands, the person who can fix it reads it. There's no game of telephone between a support agent and the engineer. Response time to "we fixed it in the app, refresh" is often under an hour. **Better product decisions.** Support conversations are the ground truth of what's broken and what's confusing. When the founder does support, product decisions incorporate that ground truth directly rather than through summaries. **What this DOESN'T scale to:** 100+ subscribers per day of support volume. At that point the model breaks; the founder would spend all day in the inbox and none of it building. Enterprise-level support expectations (SLAs, phone support, dedicated CSMs). If we ever sold to enterprise, that would be a different product with a different support model, and probably priced accordingly. 24/7 coverage. There isn't any. Emails at 2am get read the next morning. The status page is where you go if the site is down; the marketing page is honest about the support hours (roughly business hours, PST). **The escape hatches:** When support volume gets too high to handle solo, the first hire will be a support-focused person. Not "customer success" — actual support who reads the emails, answers what they can, and escalates what they can't. That's coming, not now. When response quality degrades — when the founder can't read the emails carefully anymore — the second hire is another support person. Volume triggers hire; the trigger is roughly "the last week's emails weren't handled the way the emails from month one were handled." Until either trigger, one inbox, one person. **What users can do to help:** Include specifics. "The composer isn't working" is a slower email to help with than "the composer button is disabled on the chat at /chat/abc123 as of about 3pm PT — screenshot attached, tried refresh." Send bugs to hello@ (not a support form). No triage layer, no form fields to fill out — just an email that gets read. If you're mid-conversation and need followup, reply on the same thread. Threading keeps context; new emails start over. Don't be alarmed by a same-day reply from the founder. It's the model, not urgency signaling. **What we don't do (and won't):** Automated support responses. No canned auto-reply that says "your ticket has been received" and then goes nowhere. If we can't answer within a day, you get an actual "still working on this" reply from a human. Chatbot on the marketing site. Ironic for an AI company, but our own product's whole promise is that it doesn't replace humans where humans do a better job. Support is one of those cases. Tier-based support (basic users get slower response than paid). Everyone gets the same treatment. If you're on the demo, you still get read. **The general point:** A team of one is a constraint, not a stunt. It shapes what LADLE is and isn't. Support is one of the places where the constraint shows up positively — a founder in the inbox is a feature at small scale. The trick is knowing when it stops being a feature and becomes a bottleneck. Not there yet. When we are, we'll write the next post. --- ## Reading a receipt. - Slug: reading-a-receipt - URL: https://ladle.chat/blog/reading-a-receipt - Date: 2026-07-25 - Category: NOTE - Summary: The receipts page is at /impact/reports. Every month has a document. Here's how to read one field by field — reference number, batch aggregation, per-subscriber breakdown. LADLE's core promise is a monthly receipt for the WFP donation. Every month has one. They live at /impact/reports on the marketing site. This is the field-by-field guide. **The header.** Month name and reference number. The reference number is issued by ShareTheMeal (the WFP-partner platform we donate through). Format: usually a 4-digit prefix + a letter (e.g., "4471-A"). You can look up any reference number on the ShareTheMeal side to cross-verify. **Total meals funded this month.** The count of meals that the aggregated donation from all LADLE subscribers bought in that month. This is meals funded, not meals served — WFP handles distribution on their end. The floor: this number is at least (subscriber count × 10 meals × months paid). Annual subscribers who paid in that month contribute their whole year's meal count in that month's total, so annual-heavy months look bigger. **Total USD sent to WFP.** This is the raw dollar amount transferred. It's total meals × $0.80. Small rounding for regional price variation but generally that math. **Batch number.** WFP aggregates our donations into batches for operational efficiency (a single wire transfer per batch instead of thousands of individual ones). The batch number ties our monthly total to WFP's incoming record. Same number on both sides. **Subscriber-level breakdown (optional).** Some receipts include an aggregate per-plan breakdown: how many meals came from Ladle subscribers, from Max 5x subscribers, from Max 20x subscribers, and (when there are any) from gift subscriptions. This is useful for verifying the plan mix and per-plan meal counts (10 / 50 / 100). We don't publish per-user info — that would be a privacy violation. Aggregate only. **Reference to the previous month.** Every receipt links to the prior month's receipt. Continuity is part of the trust — a missing month would be visible. **What the receipt DOESN'T include.** Distribution details — where the meals landed, which specific country, which specific program. That's WFP's operational decision, not ours. If you want to see WFP's country-level activity, their transparency reports at wfp.org are the source. Individual per-subscriber records. Your specific meals aren't personally traceable in the receipt — they're pooled into the monthly aggregate. This is intentional: aggregating protects both privacy and operational efficiency for WFP. Refunded donations. If subscribers refunded within the 7-day window in a given month, those meal donations are reversed BEFORE the batch is sent. The receipt total already accounts for reversals. **The escape hatch.** If we ever can't publish a monthly receipt on time, we say so in the changelog with the reason. If the receipt total is materially different from what the subscriber count and plan mix would predict (say, off by more than 5%), we owe an explanation. This has never happened as of publication but we've written the escape hatch policy in advance so we know how it goes if it does. **Where the receipts live.** The full archive is at /impact/reports. Every month since launch. Reference numbers are permanent — they don't rotate. If you're new to LADLE, spending five minutes reading a few months of receipts is the fastest way to verify the promise isn't marketing polish. That's the whole point of publishing them. --- ## AI for research: the four moves. - Slug: research-four-moves - URL: https://ladle.chat/blog/research-four-moves - Date: 2026-07-22 - Category: NOTE - Summary: Most 'AI for research' demos are theater. Four concrete moves that make LADLE useful in real research work — from primary sources to cited briefs. Research is a use case where AI assistants make dramatic-looking demos and mostly fail at the actual daily work. This post is what actually works, from watching researchers use LADLE. **Move one: the primary-source triage.** You have a claim you need to verify. Someone said a study showed X; a headline claimed Y; a talk mentioned Z. Instead of searching for confirmatory secondary coverage, ask LADLE (with search on): "Someone claimed [X]. Find the primary source — the actual study, report, or dataset. Give me the URL and the specific page or section number where the claim is supported. If you can't find one, tell me." The reply either gives you the primary source (good — you go read it) or admits the primary source is hard to locate (also good — that's information; either the claim is a game-of-telephone artifact or it comes from something not publicly available, and you now know which). This move alone saves researchers an hour a week that used to go to searching news coverage of the claim and getting distracted by adjacent stories. **Move two: the "what would change my mind" prompt.** You've formed a hypothesis. Before you commit to writing it up, ask: "I believe [hypothesis]. What are the strongest counterarguments? What evidence would, if true, most substantially undermine this? What are researchers who disagree usually pointing to?" The reply is a stress test. Not a rewrite, not a confirmation, not a validation — an argument against. Real researchers use this before defending a claim they care about, because the model doesn't have their ego attached to the conclusion. The best researchers we've watched use this treat the reply as a checklist: for each objection, they either address it in the write-up or note why they're setting it aside. That's the move. **Move three: the structured extraction.** You have a set of documents (papers, reports, filings) and you need to extract the same information from each. Instead of reading them all in full and taking notes: Give LADLE the schema first: "For each document, I need: author list, publication date, primary claim, sample size, key methodology, and one thing that seems weak in the design." Then feed documents one at a time (or in a batch if they fit context). The reply is a structured extraction per document. You verify the extractions against the actual documents for the ones that matter (samples the highest-stakes ones); the rest you trust with light spot-checking. This works when the extraction schema is well-defined and the documents share genre. It doesn't work well for exploratory reading where you don't know what you're looking for yet. **Move four: the cited brief.** You've done the research. Now you need to write it up for someone who won't read your full notes. Ask LADLE: "Turn my notes into a 300-word brief for [audience]. Use only claims from my notes. Cite specific documents by the identifiers I've used. If there's a claim in the notes without a citation, flag it." The reply is a brief that mirrors your notes' claims, not the model's own knowledge. The flag on uncited claims is the important part — it catches the places where you know something but haven't sourced it in your notes. This is the one that turns raw research into shareable output faster than any other move. **What NOT to do:** Don't ask LADLE to "do the research for you." The model doesn't do research; it retrieves and synthesizes. Research is about deciding what to look for, judging what you find, and reconciling conflicts. Those are the human parts. Don't trust unverified claims from the reply. LADLE hallucinates less than earlier models, but "less" isn't zero, especially on specific figures, dates, and citations. Verify anything you'd stake a reputation on. Don't use LADLE as your citation manager. It's a bad one. Zotero, Obsidian, or a real reference manager is the tool for that. **Rule of thumb.** LADLE is useful in research as a structured thinking tool with retrieval. It's not a substitute for the reading and the judgment. Used well, it saves you the mechanical work of triage, extraction, and drafting — not the intellectual work of research itself. --- ## When you should use ChatGPT instead of LADLE. - Slug: when-to-use-chatgpt - URL: https://ladle.chat/blog/when-to-use-chatgpt - Date: 2026-07-19 - Category: NOTE - Summary: Two specific use cases where the competitor is the right answer. Being honest about this is a feature, not a bug. Every comparison page on our marketing site has a "Pick them if" section that recommends the competitor for specific users. This is the version of that thinking directed at the biggest incumbent: when should you use ChatGPT instead of LADLE? Two clear cases. First: image and video generation. If your work regularly needs Sora for video, DALL·E or GPT Image for images, or the Advanced Voice mode for spoken interaction, ChatGPT Plus at $20/mo is the right tool. LADLE doesn't have those and won't for the foreseeable future — they're first-party OpenAI products with deep model integration that we can't replicate as a Claude-based reseller. If image generation is core to your day, our meal donation isn't worth trading these capabilities away for. Second: heavy dependency on Custom GPTs. If your workflow includes multiple custom-GPT integrations that other people built, or if you distribute custom GPTs yourself as part of your professional presence, the Custom GPT ecosystem is a real network you can't reproduce elsewhere. Claude's Projects work well as a personal workflow tool but aren't a marketplace. If you're in that ecosystem, staying is the right choice. We say this on the marketing site (the ChatGPT comparison page) because the alternative — pretending LADLE is universally the better product — is a lie. Honest recommendations are more useful than desperate ones. The reason we can afford to be honest: our differentiation is the meals, not the model. When we say "pick ChatGPT if you need Sora", we're not conceding anything about the meal donation. That commitment doesn't compete with Sora; they're on different axes. A subscriber who genuinely needs Sora is a subscriber who would leave anyway when they discovered we couldn't do it — better to help them decide correctly on day one than to churn them at month two. There are also cases where we think LADLE is genuinely the better call: general writing, coding, long-context reading, research with citations, and — the differentiating one — users for whom the receipt-published meal donation carries weight. We have a page arguing for that too. Both are on the site, both are true. The version of company culture we're trying to build: honest recommendations, even when they route you elsewhere. If you leave for ChatGPT and it fits you better, that's a healthier state than you staying and paying us for something that doesn't fit. Meals in that scenario go to zero, but so does the friction of a subscriber who resented us. --- ## LADLE vs Claude Team — for small teams. - Slug: ladle-vs-claude-team-for-small-teams - URL: https://ladle.chat/blog/ladle-vs-claude-team-for-small-teams - Date: 2026-07-17 - Category: NOTE - Summary: Claude Team is $25-$30/seat with shared workspaces + admin controls. LADLE is $20/seat, same model, no team features. Which one fits depends on whether you need the collaboration layer. If you're setting up an AI subscription for a small team (2-15 people), your two most likely options in the Claude ecosystem are Claude Team (Anthropic's proper team product) and LADLE (individual subscriptions billed to the company card). Here's the honest breakdown. **Claude Team.** $30/user/month monthly or $25/user/month on annual commitment. You get: shared workspaces where teammates see each other's chats, an admin dashboard for seat management, org-level data controls, SSO on the higher tiers, and priority support. Genuinely useful if teams need to share context across each other's work. **LADLE for a team.** $20/user/month, no annual commitment, no minimums. Each teammate gets an individual account. Meals accrue per subscription — a 10-seat team funds 100 meals per month, published receipt. No shared workspaces yet (see /blog/no-teams-yet for the reasoning); no admin dashboard; no SSO. **When Claude Team is the right choice.** You have a real workflow where teammates need to see each other's AI conversations — customer support with a shared knowledge base, research where two people are jointly exploring the same corpus, or any pattern where "I need to hand off the thread to my coworker" is common. The shared workspace is a real feature, not just marketing. **When LADLE is the right choice.** You have a team where each person's AI use is mostly individual — drafting, coding, research — and the "shared" part happens in Slack/Docs/Notion anyway. The $5-$10/seat/month savings adds up on a 10+ person team ($600-$1,200/year). You also get the meal donation, monthly impact report on request, and the option to hand your CFO a document showing what you funded. **Cross-pattern.** Some teams use Claude Team for the 2-3 heavy collaborative users and LADLE for the rest. Perfectly reasonable. **What's structurally different.** LADLE's per-seat $8 to WFP with a public monthly receipt. Claude Team pricing goes entirely into Anthropic's product development. Both legitimate; different values. --- ## The changelog is the marketing. - Slug: the-changelog-is-the-marketing - URL: https://ladle.chat/blog/the-changelog-is-the-marketing - Date: 2026-07-15 - Category: NOTE - Summary: Most companies write blog posts about the features they wish they had. LADLE writes changelog entries about the ones we actually shipped. Here's the specific reasoning. Look at most SaaS marketing sites. There's a blog. There are pillar pages. There are customer stories. There's an SEO effort producing "10 best X" articles. Somewhere buried under "Resources" is a changelog, updated monthly at best, three paragraphs of "we shipped some things this quarter." LADLE's marketing looks different. The blog exists (you're reading it) but the changelog is treated as a primary surface. Every ship — every real code change that a user might notice — gets a line. The receipt page is a monthly changelog for WFP donations. The founder-letter section reads like a public dev log. This isn't a style preference. It's a specific choice, and here's the reasoning. **Marketing that talks about features you don't have is a lie.** The strongest marketing in tech (for products with actual users) has a specific texture: it describes what the product IS, right now, not what the pitch deck says it will be. When companies write blog posts about their "AI-first vision" without having shipped the AI feature, users can tell. It reads as ambient BS. Changelog entries are the opposite of ambient BS. They're specific ("shipped X on date Y, replacing Z"), they're falsifiable (you can go check), and they're proof-of-work in a way marketing copy can't be. **The receipts model needs it.** LADLE's whole pitch is that we do a specific thing (send $8 to WFP each month per subscription) and prove it publicly. The receipts do the impact side. The changelog does the product side. Together they form the honest picture: here's what we did with your $20, in both directions. If the product side of that picture went dark for months while the marketing site published "How AI is transforming the way we work" content, the receipts side would start feeling like an isolated stunt. Consistent public shipping evidence is part of the trust structure. **It's easier than marketing.** Real changelog entries write themselves — the work was done, the entry is the summary. Marketing content is invented from scratch, requires positioning, requires editorial judgment, requires promotion. A three-line changelog entry is 15 minutes of work; a "10 ways to use AI for productivity" post is a full afternoon of writing content most readers won't finish. This isn't laziness. It's ROI. The changelog reaches users who already care (subscribers checking in, prospects doing due diligence, journalists writing about the space) with information they can act on. The generic marketing content reaches a broader audience with a shallower engagement, and most of that audience isn't converting anyway. **It defends against feature drift.** When the changelog is the marketing, you feel every quiet week. If nothing shipped in a week, the changelog for that week is empty, and the empty week is visible. That creates internal pressure to actually ship — small consistent progress rather than big splashy launches. The alternative (long silent periods punctuated by marketing about "coming soon" features) is how most companies drift. LADLE trying to avoid that trap. **What the changelog looks like:** Short lines with dates. What shipped, why, and a link to the affected surface if it's user-visible. Occasional deeper writeup (a blog post) for changes that need explanation — but the changelog line comes first, and the blog is the followup. Every entry corresponds to a real git commit or a real ops change. No entries invented to fill weeks; no backdated entries to make history look busier. If a week is empty, it's empty; if the changelog is quiet, the product is quiet, and that's information the reader deserves. **Where the changelog lives:** /changelog on the marketing site. Everything since launch. Public, no login. The RSS feed of it is where power users subscribe. The bottom of every marketing page could link to it and probably will eventually. That's a minor design task. **Rule of thumb.** If your marketing describes features you have, it's a changelog with a different name. If it describes features you don't have, it's a pitch deck with a different name. LADLE tries to stay on the first side of that line. --- ## AI for legal work: what we say no to. - Slug: legal-work-we-say-no-to - URL: https://ladle.chat/blog/legal-work-we-say-no-to - Date: 2026-07-11 - Category: NOTE - Summary: Legal is one of the harder use cases to draw a clean line around. Here's the specific list of legal-adjacent tasks LADLE genuinely helps with, and the ones you shouldn't use any AI for. LADLE is a general AI assistant. Lawyers and law-adjacent professionals use it, and the "AI for legal work" question comes up often. The honest answer is that the line matters more than the tool — some legal tasks are great fits for an AI assistant; others are ones no responsible AI vendor should market for. This post is the specific list. **What LADLE is actually useful for in legal-adjacent work:** **Drafting first cuts of routine documents you already understand.** NDAs, cease-and-desist letters, contractor agreements — if you know what a good version looks like and you're drafting from familiar patterns, LADLE can produce a starting draft that saves you the mechanical work. You then edit for the specifics of the current situation. **Explaining legal concepts in plain English to yourself.** If you're a founder trying to understand what an indemnification clause actually does, or a small-business owner reading a lease, LADLE gives clear explanations that are usually accurate at the "general concept" level. Not a substitute for a lawyer on your specific document — a supplement to your own reading. **Summarizing long documents you need to skim.** A 40-page RFP you need to know the shape of before deciding whether to bid. A 200-page acquisition deck. LADLE summarizes reliably at this level. You still read the parts you need to read; the summary just tells you where. **Comparing two versions of a contract.** Paste both, ask for a diff-style comparison with commentary on the changes. Useful when you're not the primary attorney but need to understand what's actually different in a redline. **Drafting internal legal memos for non-lawyer stakeholders.** Explaining a legal concept to your team, drafting policy documents, writing acceptable-use language for a product. All fine. **What LADLE (or any AI) should NOT do for legal work:** **Give legal advice on your specific situation.** This is the big one. AI produces confident-sounding legal analysis on specific factual scenarios that lawyers will tell you is wrong 30-40% of the time. The consequences of getting a legal question wrong are asymmetric — bad advice can cost you a lot more than a lawyer's fee would have. **Answer regulatory questions in regulated industries.** Healthcare, finance, insurance, education — the compliance regime is specific enough that generic AI answers are often just wrong for the jurisdiction/domain you're in. Even if the answer sounds authoritative. **Cite legal cases or statutes.** LLMs still hallucinate case citations. If a legal reply cites a case, verify the citation exists (search Westlaw / a real legal database, not a general search). LADLE doesn't have a legal-database integration; anything it cites is from its training, which means it might be paraphrasing accurately, or it might have invented a plausible-looking case name. **Represent legal analysis to a client, court, or regulator.** Whatever LADLE produces is a draft for you to edit and take responsibility for, not a document you file. Multiple lawyers have already been sanctioned in various jurisdictions for filing AI-generated briefs with hallucinated citations. **Contract review as the primary review.** Use LADLE to catch obvious issues before your lawyer sees it (saves them time = saves you fees). Do NOT use it INSTEAD of your lawyer for anything material. **Why we're careful to say this out loud:** LADLE's marketing doesn't push the "AI for legal work" angle because we don't think it's a responsible pitch. The line between "helpful legal-adjacent tool" and "unauthorized practice of law" is fuzzier than most SaaS categories, and vendors who blur it push liability onto their customers. **A concrete recommendation for LADLE-using lawyers and law-adjacent professionals:** Use LADLE for the writing and organizational work you'd do anyway (drafts, explanations, summaries). Get your actual legal advice from actual lawyers, ideally ones who know your specific situation. LADLE at $20/month doesn't replace the $500/hour attorney, and pretending otherwise is where the trouble starts. Rule of thumb: if the stakes of being wrong are more than the cost of a lawyer's opinion, get the lawyer. --- ## The temporary chat: what it is, and what it isn't. - Slug: the-temporary-chat - URL: https://ladle.chat/blog/the-temporary-chat - Date: 2026-07-08 - Category: NOTE - Summary: Temporary chats never write to the database. The Anthropic API still processes the request. Here's the specific difference, and what temporary is (and isn't) for. LADLE has a temporary chat mode. You toggle it from the composer. The reply streams normally. Nothing about the request or the reply gets written to the LADLE database. Nothing shows up in your history. Refresh the page and the chat is gone. Two questions come up about this feature. Let me answer both directly. **"Is the request still sent to Anthropic?"** Yes. The model produces the reply through the same Anthropic API call as a normal chat. Anthropic's stated policy is that API requests aren't used for training, and their retention window for API traffic is 30 days by default (configurable to zero for enterprise). We haven't invented a way to make a request to a model without the request being made. So "temporary" at LADLE means: LADLE doesn't store it. It does NOT mean: the request never happened at any layer of the stack. The Anthropic API sees it either way; it just isn't logged in our database or associated with your account beyond what Anthropic already sees. **"What is temporary chat FOR, then?"** A few practical uses: Testing a prompt without cluttering your history. If you're iterating on a prompt to see what shape works, you don't necessarily want six half-attempts in your recent chats. Temporary chats let you experiment without leaving evidence in your own workspace. Sensitive drafts you don't want to accidentally re-open on the wrong device. A resignation letter, a medical question, a break-up message. Not "we've protected the API from processing it," but "no one who picks up your laptop later can scroll through your history to it." Questions you'd delete anyway. If your answer is "I'm just going to delete this after," temporary is the delete-in-advance version. One less step. **What temporary chats DON'T do:** They don't stop the request from being processed by Anthropic. They don't provide extra encryption. They don't route through a separate provider. They don't skip the usage counter — the tokens still count against your monthly budget. They don't skip the meal accrual either — you funded meals; your usage funded meals. **What we deliberately kept:** The Anthropic policy that governs LADLE API traffic is the same for temporary and persistent chats: no training on API data, standard 30-day trust and safety window. If Anthropic ever changed that, we would tell you. The kitchen panel works on temporary chats. You can still see which model produced the reply and what the routing decision was. That's local information — it lives in the response headers, not in a database. **A subtle point.** Temporary is not "incognito." Incognito browser modes hide activity from the browser's local storage but do nothing about what your ISP or the sites you visit see. Temporary at LADLE hides activity from LADLE's database but does nothing about what Anthropic's API sees. If your threat model needs "not even Anthropic sees it," temporary chat isn't the tool. Running an open-weights model locally is. That's a different product than LADLE, and we're happy to recommend it for the cases where it matters. For most everyday cases where you don't want the chat sitting in your history? Temporary is exactly what it should be. --- ## LADLE vs Perplexity — for citation-heavy work. - Slug: ladle-vs-perplexity-for-citations - URL: https://ladle.chat/blog/ladle-vs-perplexity-for-citations - Date: 2026-07-03 - Category: NOTE - Summary: Perplexity is a citation-first answer engine. LADLE is a chat assistant with citations when search is on. Different tools for different jobs. Perplexity built its whole product around inline citations. Every answer is annotated with source links, and the primary interface is a search-style query box. LADLE is a chat assistant that cites when web search fires. These are meaningfully different product patterns. **Perplexity's strengths.** Citation-per-claim density is higher. The UI is optimized for verifying every sentence against a source. For "answer this question with sources I can hand to my editor," Perplexity is a good default. The Pro tier costs $20/mo. **LADLE's strengths.** Multi-turn conversation is smoother — LADLE remembers what you asked three turns ago, Perplexity's chat mode is thinner. For extended research where you're refining questions across a session, chat structure wins. Also: Claude Sonnet is a stronger writer than most search-first models, so if the OUTPUT of your research is prose, LADLE's replies are usable as first drafts more often. **When to pick Perplexity.** Fact-checking a specific claim, source-hunting for a specific citation, quick research where citation-density > conversational depth, competitive research where you need to trace every claim to a URL. **When to pick LADLE.** Multi-turn research sessions, research that ends in a written deliverable, working through complex reasoning where citations matter but aren't the primary output, or when you want the meal-donation social contract in your $20/mo AI spend. **Can you use both?** Yes. A common pattern: Perplexity for the initial fact-hunt (find the sources), LADLE for the synthesis (turn the sources into a written brief). $40/mo total, and they don't overlap much. **What's actually different structurally.** LADLE's $20 sends $8 to WFP with a public receipt. Perplexity's $20 stays entirely within Perplexity's business. If the structural commitment matters, LADLE. If not, they're both solid — different products. --- ## Building LADLE with a team of one and a lot of Claude. - Slug: building-ladle-solo - URL: https://ladle.chat/blog/building-ladle-solo - Date: 2026-07-02 - Category: NOTE - Summary: The specific stack that lets one person ship this product, the specific leverage the model gives, and the parts of running a real SaaS that don't automate. LADLE is built and operated by one person, so far. Not because that's a moral commitment; because at current scale the marginal engineer costs more per month than we can absorb without cutting the meal donation. When the math flips, this changes. Here's what the current stack looks like and what actually works. The stack: Next.js 16 on Vercel, Postgres on Neon (later this year, self-hosted), Stripe for billing, ShareTheMeal API for meal donations, Resend for email. Claude Sonnet via Anthropic API for inference. That is the whole production stack. Zero analytics vendors, zero marketing automation, one auth library (Lucia), one component library (none — everything is hand-rolled with CSS variables from a DESIGN.md source of truth). Where Claude Sonnet is the specific leverage: (1) writing the copy for every content page on the marketing site, with heavy editing rather than acceptance-of-drafts; (2) reading unfamiliar code in Anthropic's SDK before I extend it; (3) debugging Stripe webhook edge cases by pasting logs and stack traces; (4) drafting the honest cancel flow copy and the trust page; (5) drafting every FAQ, comparison entry, and help article on the site before I edit them. Where the model is not the leverage: business decisions, pricing decisions, the choice of what to build vs cut, customer conversations that require judgment about who I'm talking to, the actual reconciliation logic for the ledger (I wrote and reviewed every line of that myself), incident response. The pattern that scales: I use the model to draft, then I edit hard. First drafts from Claude land at maybe 60-70% of shippable quality. The last 30-40% is where the voice lives, and the model can't do that without a specific-enough prompt that it's easier for me to write it. The parts that still take a specific human: keeping the operating ledger balanced, filing quarterly taxes, having the payment failure conversation with the one card issuer that keeps declining our merchant category code, responding to the four support emails a day that require actual reading (out of maybe fifteen total). Also every legal document. Also every subscribed-but-doesn't-know-what-to-ask user who deserves a personal reply. The scale question I get asked: at what point do you hire? My honest answer: when the growth rate is fast enough that a specific unshipped feature is losing us more meals per month than a mid-senior engineer's salary would cost. That crossover isn't close. When it is, this page will change. Nothing about running LADLE this way is heroic. It's a specific set of trade-offs that fit this specific business at this specific scale. Very much not the pattern I'd recommend for a different kind of company. --- ## What happens when Anthropic drops prices. - Slug: what-changes-when-anthropic-drops-prices - URL: https://ladle.chat/blog/what-changes-when-anthropic-drops-prices - Date: 2026-07-01 - Category: NOTE - Summary: Anthropic has cut Claude API prices twice since we launched. The subscription stayed $20 both times, the meals stayed at ten, and the headroom got passed through as more usage. Here's the reasoning. Anthropic has cut Claude API prices twice since LADLE opened for subscriptions. Both cuts were meaningful (roughly 20% each time on the relevant token classes). This post is the explicit accounting of what we did with the extra room. **What we didn't do:** We didn't drop the subscription price. $20/month is the price. Dropping to $18 would have been a natural pass-through, but it would have opened a real design question about the meals ($8 becomes an awkward fraction of a smaller price, and the marketing anchor loses its round-number simplicity). We chose to keep the price and pass the value differently. We didn't increase the meal donation. Raising to $9 per subscription would have been the flashy move, but it changes the marketing (10 meals becomes 11-12), the receipts have to explain the change, and if Anthropic ever RAISES prices we'd be caught having to lower the meal count publicly. The floor of 10 meals is a load-bearing constraint; we don't want to move it upward without confidence it can stay there. **What we did:** We increased the per-subscriber usage headroom. In practice: the monthly compute budget baked into each plan went up (Ladle from ~$8/month of API to ~$9.20; Max 5x from ~$45 to ~$47; Max 20x similarly). Users who were regularly bumping into their monthly compute cap now have more room before they hit it. Users who never came close saw no change. **Why this way:** The subscribers who most benefit from a price drop are heavy users. They're the ones for whom the compute budget was the constraint, not the meals. Giving them more headroom is a real product improvement for the people who feel the constraint, without touching the marketing narrative that everyone else counts on. The subscribers who don't benefit — light users, casual users, users for whom the meals are the primary reason to subscribe — get the same product they were already using, at the same price, funding the same number of meals. Nothing worse for them; nothing better either. **What we changed in the pricing page copy:** Barely anything. The plan pages still say "Ladle base includes generous daily use" without pinning a specific token count, because pinning a specific count would create weekly-updated marketing every time the model changed. The internal budgets adjusted; the external promise stayed steady. We did add a line to the "how we spend the $20" breakdown: "if inference costs drop, that flows through as usage headroom, not as a lower price." Honest, and it explains why the price is stable even as costs change. **When would this reasoning change?** If Anthropic dropped prices by 50%+ in a single move, the "pass-through as headroom" strategy would produce more headroom than any reasonable subscriber could use. At that point we'd revisit the meal donation floor — a $12 or $15 meal portion of the $20 might be feasible without breaking the marketing. But the change would be structural, announced deliberately, and permanent. If Anthropic RAISED prices by 20%+, we'd absorb the first cut from margin ($3.50/subscriber has some room). Beyond that we'd have to raise the subscription price and be honest about why. Meals wouldn't drop — the meal floor is untouchable. Ops margin would compress first; price would rise second; meal count would never move down. **The general principle:** Price stability is a feature. Meal count stability is a feature. Both matter more to the LADLE story than optimizing every dollar. Passing cost savings through as capacity (rather than price) preserves both while still giving heavy users the win. If Anthropic drops prices again — and they probably will as models get cheaper to serve — expect the same playbook. Same price, same meals, more room. --- ## June 2026 changelog roundup. - Slug: changelog-roundup-2026-06 - URL: https://ladle.chat/blog/changelog-roundup-2026-06 - Date: 2026-06-30 - Category: CHANGELOG - Summary: Everything that shipped in June, in one post. Streaming latency cuts, the My Table launch, project imports, and one regression we owned publicly. Monthly changelog roundup. Everything that actually changed, one bullet each. **Shipped** - **My Table screen** — personal impact ledger inside the product. Every subscriber can now see the exact plates their subscription has funded, month-over-month, with the reference number linking to the public ledger. - **Streaming latency improvements** — first-token latency dropped from ~450ms average to ~380ms after a transport tweak on our side (context caching enabled by default in production). - **Project imports** — you can now import a Project from another LADLE account's export. Useful for teams sharing voice guides and reference documents. - **/apps status page** — the previously-vague apps roadmap is now a specific status table. iOS in TestFlight, Android/Mac/Windows landing next. - **/gift flow** — you can gift 1, 3, 6, or 12 months of LADLE. Meals land in the recipient's name. **Changed** - **Composer chip labels** — "Search" chip renamed from "Web" for clarity. Purely cosmetic; behavior is unchanged. - **File upload limit clarification** — the 30-attachment-per-chat limit is now shown as a counter in the composer when you're approaching it. Previously it was silent until you hit it. - **Onboarding** — new subscribers now see a two-step orientation on first login (previously landed directly in the chat interface). Skippable. **Fixed** - Streaming responses would occasionally lose the last few tokens when the connection tab was backgrounded on iOS Safari. Reproduced, fixed. - The nav odometer would stop ticking after 4 hours of continuous scroll on the same page. Client-side timer bug, resolved. - The "you've been signed out" flow triggered incorrectly for users on airplane wifi with dropped-packet reconnects. Made more forgiving. **Regressed (and owned)** - The 2026-06-14 deploy of My Table shipped with a bug where the plate-grid animation played on every page load, not just on first-load or new-meal events. Multiple subscribers reported the visual noise. Fixed within 6 hours. Owning it here because we should have caught it in QA and didn't. **Next month's queue (not commitments — just the priority order)** - iOS TestFlight expansion. - Sitemap and llms.txt for the site content shipped in Q2. - Dynamic OG images. - The first-cut of an accessibility audit. If you use LADLE and any of the shipped changes affects you in a specific way we should know about, email support. Same-day response. --- ## A note on the Daylight theme. - Slug: a-note-on-the-daylight-theme - URL: https://ladle.chat/blog/a-note-on-the-daylight-theme - Date: 2026-06-27 - Category: NOTE - Summary: Most 'light mode' in software is 'dark mode with the values inverted.' LADLE's Daylight is a different room, not an inversion. Here's the design thinking. LADLE ships with two themes. Midnight is warm near-black with an umber undertone; Daylight is warm cream with a coral accent. Users can toggle in Settings or in the onboarding step. New users default to Daylight. Most software's "light mode" is just the dark theme's colors inverted — background swaps from black to white, text swaps from white to black, done. LADLE's Daylight is different in a way that took real work. This post is why. **The problem with inverted dark mode:** When you invert a dark theme, you preserve the accent choices (colors, shadow depths, animation timings) that were tuned for a dark canvas. Amber on black is warm; amber on white is sallow. Shadow-based depth reads clearly on dark; shadows on light need to be lighter, softer, and more diffused or they feel cheap. Type sizes optimized for high-contrast dark reading are too big on a light canvas where the eye has more room to breathe. The result: most software's light mode feels like a stepchild. Users prefer dark not always because it's better but because the light version is worse. **What we did for Daylight:** Different palette. The Daylight canvas is #F5EFE1 (warm cream), not white. Text is #191413 (warm near-black), not pure black. Accent is Claude-coral (#CC5A2C), not the amber that lives on Midnight. Every color was re-picked for the light context, not inverted. Different depth model. Midnight uses tone + inner-light glow to create depth (surfaces get slightly warmer as they lift). Daylight uses soft ink shadow + warmth (surfaces get slightly whiter and lift with a light shadow). Both feel dimensional; neither is the other with colors flipped. Different typography weight. Same fonts, but headings are ~500 weight on Daylight vs 500 on Midnight for the same visual density — pure black text on cream reads heavier than warm ivory on warm dark at the same weight. Different accent behavior. On Midnight, the Claude-amber pulses softly to draw attention (it's the "warmth in the candlelight" affordance). On Daylight, the same role is played by Claude-coral but with a different rhythm — coral on cream is bolder than amber on dark, so it doesn't need to pulse as often. **The reference:** We were explicit about the reference model: Claude.ai's cream theme, which is the best consumer light-mode design in the AI product category. Not because we're copying — the values, spacing, typography are ours — but because "make cream feel as considered as Anthropic did" was the north star. **What Daylight is for:** Reading. Long-form drafts, research, documentation, review — anything where you're going to spend an hour looking at text — reads better on cream than on dark. This is empirically true and it's why every book, magazine, and long-form website defaults to light. Working during the day. If your monitor is next to a window, Midnight is fighting your ambient light. Daylight isn't. Users who prefer light mode in general. Roughly 30% of users prefer light; another 40% prefer dark; the rest use system-follows-time. All three options should be considered surfaces, not afterthoughts. **What Midnight is for:** Focus. Dark themes reduce visual noise from surrounding elements; the eye is drawn to the text. Coding, deep thinking, evening work — Midnight is the room. Signature aesthetic. The dark warm-black-with-amber is LADLE's brand — every marketing surface uses Midnight. It's what we look like when someone asks "what does LADLE look like." **One thing we deliberately avoided:** The "auto-switch based on time of day" default. System-follows-OS is available, but we don't force switch on wall-clock time. Users tend to prefer stability over cleverness — waking up in the morning and having your workspace look different than yesterday is jarring in a way that's not worth the theoretical eye benefit. **A specific rule for the design:** Marketing pages always stay Midnight. Even for users who prefer Daylight in the app, the marketing site keeps its brand consistency. That was a deliberate call — the app is the room you work in; the marketing is the poster on the wall. They can have different palettes and both feel right. If you're using Daylight and something looks off — text that blends in, chip that's hard to see, hover state that doesn't register — write in. Daylight is younger than Midnight and getting continually tuned as we notice things. --- ## The 200K context window: what actually changes. - Slug: two-hundred-k-context-what-changes - URL: https://ladle.chat/blog/two-hundred-k-context-what-changes - Date: 2026-06-25 - Category: NOTE - Summary: 200K tokens is about 500 pages of dense text or a full novel. Here's what that actually enables — and where the practical ceiling still lands, because 'the model can read 500 pages' isn't the same as 'reads them well.' Claude Sonnet's context window is 200,000 tokens. In plain terms: about 500 pages of dense text, or a full novel, or ~15 hours of chat history, or a fair-sized codebase, or two 10-K filings plus room for questions. That's a big number, and it's a real technical capability, and it also comes with practical caveats that don't fit on a spec sheet. This post is the honest guide. **What the number means.** A token is roughly 3-4 characters of English, so 200K tokens is around 700K characters or 150K words. A typical book chapter is 5-10K words. So a 200K context window fits about 15-30 chapters, plus your prompt, plus room for the model's reply. **What the number enables that shorter windows couldn't.** - Attach the full text of a long PDF and ask cross-referencing questions across it - Paste a large codebase (or several files) and ask for a review that considers the whole shape - Keep a long research chat going without losing what you discussed 40 turns ago - Compare two long documents side-by-side without summarizing either one first - Give the model a lot of your own writing as context for tone matching **The practical caveat: attention doesn't scale evenly.** The model can technically see all 200K tokens. In practice, its attention on any specific fact deep in the middle of a long context is weaker than its attention on facts at the beginning or end. This is a well-known finding — "lost in the middle" — that applies to every long-context LLM, not just Claude. It means: for critical information, either put it near the top or the bottom of your context, or ask about it directly rather than assuming the model will surface it unprompted. **The practical caveat: cost.** Long context is billed per token. A 200K-token turn costs meaningfully more than a 10K-token turn. LADLE's compute budget is a monthly cap; heavy long-context usage will move you toward that cap faster than short-context usage. Nothing hidden about this — it's just physics. **The practical caveat: latency.** Longer context = longer time to first token. A 100K-token attachment adds several seconds to the "thinking about the input" phase before the reply starts streaming. Not a lot, but noticeable in a conversation. **What we do about this at LADLE.** The routing rules default to short-context patterns for most turns — casual chat, drafting, short questions — where a big window would just cost more without helping. When you attach a large file or paste a long document, the router shifts to the full-context mode automatically. You don't need to think about it. Chat history in LADLE also lives in the window: your last N turns are re-sent with every prompt so the model has continuity. We cap that at ~50 turns or ~200K chars, whichever is smaller, so a very long chat doesn't blow the budget on every reply. **The rule of thumb.** 200K is generous, not infinite. Treat it as a resource: use it when a big context is genuinely what the task needs (a long PDF, a codebase review, a document comparison), and don't chase it as a feature for its own sake. --- ## Why we don't sell teams (yet). - Slug: no-teams-yet - URL: https://ladle.chat/blog/no-teams-yet - Date: 2026-06-20 - Category: NOTE - Summary: You can subscribe your team to LADLE today — same $20/seat price. But there's no shared workspace, no admin dashboard, no SSO. That's deliberate for now. Here's the reasoning. LADLE has a business page. It says: same $20/seat as consumer, no annual commitment, monthly impact reports on request. That's true, and companies do subscribe their teams that way today. What we DON'T have yet is a proper team product. No shared workspace where colleagues can see each other's chats. No admin dashboard for a team lead to manage seats. No SSO integration with Okta or Google Workspace. No org-level data controls beyond what individual users already have. Companies who subscribe today set up individual accounts for each teammate, expense the seats on a company card, and use LADLE as N-individuals-with-shared-billing rather than N-users-of-a-collaborative-product. Here's why we haven't built more. **Reason one: the collaboration features that matter are hard to get right.** Shared workspaces make sense when there's a workflow where two people naturally need to see each other's context. Most AI assistant usage isn't like that. Writing an email doesn't need collaboration. Reviewing code doesn't need collaboration. Research doesn't need collaboration until you're writing the report, at which point Google Docs is a better collaboration surface than a chat product would be. We haven't seen a compelling case for "our team needs to share this AI chat" that Slack + a shared doc doesn't solve better. **Reason two: admin dashboards are surface area we'd have to maintain forever.** A proper team admin dashboard means seat management, invite flows, role permissions, per-user usage caps, org-level SSO, audit logs, data export controls at the org level. Every one of those features is a small forever-cost of maintenance and support. We'd want them to serve a real need, not to check a box for procurement. **Reason three: LADLE's value prop is unusually simple, and team features can complicate the pitch.** "One AI subscription. $8 of every seat funds meals through the WFP. Same product for solo users and companies." Adding a team tier introduces "well but the team version has..." into every conversation. We're not in a hurry to complicate that. **What would change our minds:** If more than a small handful of businesses tell us their procurement or IT process actually needs SSO or an admin dashboard to say yes, we'll build the specific piece that unblocks them. Not the full team suite; just what's needed. If we get repeated asks for a specific collaboration primitive (shared prompts library across a team, for example), that specific thing is probably a good build. **What we already do that helps teams:** The monthly impact report — PDF, CSV, or JSON — is available to any subscriber. If your team is subscribed and the CFO or CSR lead needs a report of what the team funded to WFP this month, we generate it on request. That's already built and available now. The business page also has a procurement Q&A: DPA availability, subprocessor list, data flow diagram, security incident history. Everything a vendor-review team asks for in the first round. So — no shared workspaces yet. If your team is buying seats, it works today. If your team needs specific team features and can't move without them, tell us which ones. The list of what we build next is driven by what you actually ask for. --- ## LADLE vs ChatGPT — for research work. - Slug: ladle-vs-chatgpt-for-research - URL: https://ladle.chat/blog/ladle-vs-chatgpt-for-research - Date: 2026-06-19 - Category: NOTE - Summary: For research work, the two products differ on citation style, long-context behavior, and search integration. Neither wins outright — the choice depends on your specific workflow. Researchers spend a disproportionate amount of time in AI assistants, so the small quality-of-life differences matter more here than for casual use. Here's the honest comparison. **Citations.** LADLE runs on Claude Sonnet with Anthropic's server-side web search tool. When search fires, the reply includes footnote-style citations with real URLs you can click to verify. Modern ChatGPT also cites, with similar quality. Both hallucinate citations occasionally with search OFF — always verify anything you'd stake reputation on. **Long-context recall.** Both support 200K+ token contexts. Both have the "lost in the middle" problem — attention on facts in the middle of long inputs is weaker than at the beginning and end. For research work with long papers, put your specific question near the top OR the bottom of the context and don't assume the model will surface middle-of-document facts unprompted. **File handling.** Both accept PDFs. LADLE caps at 30 attachments per chat with a 100MB hard limit per file; ChatGPT's limits are similar. For a research session across 20+ papers, both work; both benefit from focused chats rather than one enormous chat. **Extended thinking.** Both models offer a "think longer" mode for complex reasoning. LADLE surfaces this as a Deep-mode composer toggle. ChatGPT has a similar feature (o1-style). For multi-step research reasoning, either helps meaningfully. **Speed.** Roughly equivalent on typical research turns. Sonnet is a hair faster on shorter contexts, GPT slightly faster on very long contexts. Not decisive either way. **What's actually different.** The meal donation ($8 of every LADLE subscription to WFP with a monthly receipt) doesn't help your research, but it is the whole reason LADLE exists as a separate product from Claude.ai and ChatGPT. If that structural commitment matters to you, LADLE. If not, use whichever chat surface you prefer. **Recommendation.** If you have workflows tuned in ChatGPT already, switching costs are real — stay put unless you specifically want the LADLE positioning. If you're new, either is a strong research assistant; pick on chat surface preference and social contract fit. --- ## The meals map — where they land, and why we show it. - Slug: the-meals-map - URL: https://ladle.chat/blog/the-meals-map - Date: 2026-06-16 - Category: NOTE - Summary: The map at /meals-map shows WFP's active operations. Meals we fund land somewhere on this map — but we don't pretend to say 'your specific meal fed this specific person.' LADLE has a page at /meals-map that shows the countries where the WFP is currently operating. Marker size roughly scales with the operational footprint. The intent is to give subscribers a geographic sense of where their contribution lands. This post is about what the map shows, what it deliberately doesn't, and why we don't overreach on the attribution question. **What the map shows.** WFP operates in ~120 countries. The map shows the current active operations pulled from WFP's public operational data. Bigger markers = larger operational presence in that country (measured roughly by number of people reached). Hovering shows the country name and the program type (school feeding, emergency response, resilience programming, etc.). The data isn't real-time — it updates as WFP publishes updates, typically monthly. What you see is a recent snapshot, not a live feed of "meals being served right now." **What the map DOESN'T show, and why.** Individual attribution. We can't say "your $8 in June funded three meals in Yemen and two in Ethiopia." The WFP donation model doesn't work that way. Your donation is pooled into the aggregate LADLE contribution, then into the larger WFP operational budget, then allocated by WFP's operational planning across programs and geographies based on need. Pretending otherwise would be marketing fiction. "You fed a specific child in a specific place" is emotionally powerful but factually not how humanitarian food aid works. WFP allocates by need, urgency, and logistics — not by donor attribution. **Why we show the map anyway.** Two reasons. One: subscribers reasonably want to know where their money geography-goes, even if not person-specifically. The map answers that: WFP works in these countries, at these approximate scales. Your contribution lands somewhere on this map, weighted approximately by their active operational allocation. Two: it's a check against the abstract. "$8 buys meals for people who need them" is true but distant. Seeing that WFP is currently running school feeding in 60+ countries, emergency response in dozens more, resilience programming across multiple continents — that grounds the abstraction in specific places. **A subtlety about "where the meal is served."** There's a meaningful gap between where WFP receives donations (mostly rich countries) and where WFP serves meals (mostly countries in acute humanitarian need). This gap is intentional — humanitarian aid moves resources from where they can be raised to where they're needed. If we ONLY showed the countries where meals are served, someone might wonder why "their" meals appear in places they haven't heard of. Because that's where the need is. **What we would add if we had more data:** Per-region breakdown of WFP's monthly disbursement, so subscribers could see the specific mix of countries their donation batch supported. WFP doesn't publish this at monthly granularity through public APIs. If they do, we'll add it. Program-type breakdown (emergency vs. school feeding vs. resilience). Same story — data isn't consistently public. Would help contextualize the impact. **Where the map is now.** It's a static-ish view sourced from WFP's operational updates, refreshed when they publish new data. Good enough to answer "where does WFP work" and "where do meals I've funded land, geographically." Not enough to answer "which meal was mine." And, as above, that second question doesn't have an honest answer anyone can give. If you want the deeper WFP transparency reports (financial breakdowns, per-country allocations, program strategy documents), wfp.org is the source. The LADLE meals map is a summary. WFP's own site is the depth. --- ## AI for teachers: the honest use cases. - Slug: teachers-honest-use-cases - URL: https://ladle.chat/blog/teachers-honest-use-cases - Date: 2026-06-13 - Category: NOTE - Summary: AI in the classroom is either a panic or a promise. The truth is smaller and more useful — a specific list of what LADLE helps teachers with and what it can't. Teachers are one of the largest professional groups using LADLE. Their feedback has been unusually specific about what works and what doesn't, so this post exists to compile that experience. **What LADLE helps teachers with, honestly:** **Differentiation drafts.** You have the same content to teach and three reading levels in the class. LADLE will produce three versions of a short explanation at three different reading levels in a minute. You edit them for accuracy and voice, but the mechanical work of "restate this concept at grade 6, grade 8, grade 10" is exactly what the model is good at. **Rubric generation.** Give it the assignment, the learning objectives, and the grade level. Ask for a scoring rubric with 4 levels. You'll get a starting draft that's 80% right; the last 20% is you knowing your specific class and standards. This saves an hour on the mechanical rubric work. **Parent communication drafts.** "Draft a note to a parent about a student who's struggling with X. Be direct, kind, and specific. Suggest one concrete thing we could try together." The output is a starting draft; you edit. Faster than starting blank, and often catches phrasing better than the tired-Friday version of you. **Lesson-plan variation.** You've taught the same unit for five years. It's tired. Ask LADLE for three different ways to open the unit that hit the same objectives but through different angles. You'll get some useful reframes and some that don't fit your class — the useful ones justify the exercise. **Feedback for student drafts (not grading).** Paste a student's essay draft (with permission and appropriate anonymization). Ask for observational feedback focused on structure and reasoning, not grading. The reply is a starting point for you to refine into feedback you give — it's not the feedback itself. Students should hear your voice, not the model's. **What LADLE does NOT do well for teachers:** **Grading.** Even with a rubric, LADLE grading isn't reliable enough to trust for stakes. Two problems: (1) the reasoning behind a grade is often subtle and rubric-inconsistent in ways only a human notices, and (2) students deserve grades from their teacher, not from a machine. Use it for feedback drafts, not for grades that go on records. **Answer keys for math or science with high accuracy requirements.** For anything past basic arithmetic, verify answers manually. LLMs still make specific errors on multi-step math and physics that a student catching would embarrass everyone. **Detecting AI-written work.** Every "AI detector" has high false-positive rates. LADLE can't reliably tell you if a student's essay was AI-written. Don't use it that way; you'll wrongly accuse honest students. **Substituting for your subject knowledge.** LADLE will confidently produce content in your subject area that sounds correct and contains subtle errors only you would catch. Use it for phrasing help and structural drafts, not for content you don't know deeply enough to verify. **On the "students are using AI to cheat" question:** The reality is: some are. Some will keep doing it. AI writing detection isn't reliable. The pedagogical response is to rethink what assignments actually assess, not to try to detect AI harder. Assignments that can be completed by pasting the prompt into an AI shouldn't be your primary assessment. Oral defenses, in-class writing, project-based work, and process artifacts (drafts with revision history) are harder to shortcut. That's a whole other conversation and not really a LADLE conversation. But if you're a teacher weighing whether to teach WITH AI or against it, the middle path (teach with it, redesign assessments so it doesn't undermine learning) is where most of the effective teachers we've talked to have landed. **Rule of thumb.** Use LADLE for the mechanical work you'd otherwise do at 10pm on Sunday. Save your time for the parts of teaching that only you can do — knowing your students, giving your feedback, holding the room. AI helps with the drafts, not with the teaching itself. --- ## How the meal ledger actually works. - Slug: how-the-meal-ledger-works - URL: https://ladle.chat/blog/how-the-meal-ledger-works - Date: 2026-06-11 - Category: NOTE - Summary: The specific engineering of the meal donation: two ledgers, one monthly transfer, one reference number per month, published the day it clears. The meal ledger is deliberately simple, both because it's easier to reason about and because complexity is where trust erodes. Here's how it works, top to bottom. We run two accounting ledgers separately. The operating ledger tracks the $12 per subscription that funds Anthropic API costs plus our overhead. The meal-committed ledger tracks the $8 earmarked for WFP. These are literally separate database tables with separate transactional boundaries. Money can flow into either from a Stripe charge; money can't flow between them. When Stripe processes a successful $20 charge, our webhook handler writes two rows: $12 to operating, $8 to meal-committed. Both are timestamped with the exact moment of the charge. If the charge fails, no rows are written. If the charge is refunded (rare, but happens), both rows are reversed atomically in the same transaction. On the last calendar day of each month, a scheduled job sums the meal-committed ledger for the month and initiates one aggregate donation to ShareTheMeal via their API. The transfer typically clears within 30 minutes, during which the job is in a "pending" state. On success, ShareTheMeal returns a unique reference number that we associate with every meal-committed row for that month. The receipt is published as soon as the reference number is confirmed. It appears in three places atomically (same database transaction): the public Impact page ledger, the current month's Impact Report, and every subscriber's personal My Table screen where the plates for that month "confirm" from pending to committed status. Reconciliation checks run continuously. Every 15 minutes, a job verifies that: - The sum of the meal-committed ledger equals the sum of all successful $20 charges times $8 (adjusted for refunds). - The sum of published receipts equals the sum of meal-committed rows tagged with a reference number. - Every reference number on the public ledger has a corresponding response record from ShareTheMeal. Any drift triggers a page. In four months of operation, we've had zero true drift; two false positives from clock-skew between our systems and Stripe (resolved within minutes, published on the status page). The intentional simplifications: we do not compound meal counts across time. If a subscriber's charge fails and they re-subscribe later, their historical meals stay counted at their original values — we don't retroactively adjust anything about past batches. And we do not partial-donate. Every month's transfer is a single all-in donation; if for any reason we couldn't execute it (bank holiday, ShareTheMeal outage, our own bug), the whole batch waits until the next successful attempt, with the status page reflecting the delay. Everything on this page is verifiable independently. The reference numbers on the ledger are the specific bookkeeping identifiers ShareTheMeal uses; you can email their donor support and they'll confirm the amounts and dates. --- ## What we won't pretend AI can do. - Slug: were-honest-about-what-ai-cant-do - URL: https://ladle.chat/blog/were-honest-about-what-ai-cant-do - Date: 2026-06-08 - Category: NOTE - Summary: A specific list of tasks LADLE (or any current AI assistant) is not good at. From the vendor whose interest is you overestimating it. Every AI vendor's marketing implies the model can do more than it can. That includes ours, honestly — every "AI assistant that helps you write, code, and think" tagline glosses over the failure modes. This post is the specific list of things Claude Sonnet (the model LADLE runs on) is bad at, from the perspective of a vendor that would financially benefit from you not knowing them. **Arithmetic and math beyond the trivial.** Sonnet handles addition and small multiplication reliably. It fails on multi-step math with confident-sounding wrong answers. If your task involves numbers where precision matters, either use extended thinking (which helps meaningfully) or double-check the arithmetic in a calculator. Do not paste a spreadsheet and ask "what's the total" without verifying. **Citations and references.** Sonnet will produce plausibly-formatted citations that don't exist. This includes: fake academic papers, fake court cases, fake URLs, fake book titles. Web search (when enabled) reduces this to near-zero for anything the search finds. Without search, treat every specific reference as unverified until you check. **Recent events.** The training data has a cutoff. Anything after that cutoff, the model doesn't know unless web search is on. If you ask about something recent and search is off, you'll get old information or a plausible guess. **Multi-step reasoning with tight verification requirements.** Complex logic puzzles, multi-step math proofs, long agentic workflows where each step depends on the previous — Sonnet is competent but not reliable. If the answer needs to be right, verify each step. Extended thinking helps but doesn't eliminate. **Long-context recall of specific facts.** With a 200K context window, Sonnet CAN see everything you gave it. It doesn't always ATTEND to the middle of long inputs as strongly as the beginning and end (the "lost in the middle" phenomenon that affects every LLM). Put critical facts near the top or bottom, or ask about them specifically. **Following complex multi-step instructions perfectly.** Give Sonnet 15 rules to apply to a document. It will get 12-14 right and quietly ignore or misinterpret 1-3. Verify against your rules; don't trust the "here's the output" reply blindly. **Producing consistent output at high volume.** If you ask for "20 variations of X," you'll get 20 things, but they'll cluster in structure and phrasing. LLM diversity is narrower than the count suggests. If you need actual variety, generate in smaller batches with different prompts. **Legal advice on your specific situation.** LLM legal analysis is confidently wrong 30-40% of the time on specific factual scenarios. Use it for concepts and drafts; get actual lawyers for actual advice. **Medical diagnosis or treatment recommendations.** Same failure mode. Sonnet will produce a confident-sounding differential that's a poor substitute for a doctor who has your history. Don't rely on it for anything material. **Handling ambiguity gracefully.** Faced with an ambiguous question, Sonnet often picks an interpretation and runs with it rather than asking. The pick is sometimes right, sometimes wrong. Being explicit in your prompts reduces this. **Generating text that meets specific length constraints.** "Write exactly 100 words" is not reliably produced. You'll get 80-140. If precise length matters, generate and trim. **Judging its own confidence.** Sonnet doesn't reliably know when it's guessing. It says "I'm confident" and "I'm not sure" but these are more style choices than calibrated probability. Verify anything the confidence-language suggests you should trust. **What we're NOT saying:** We're not saying LADLE is bad. It's a genuinely useful daily assistant for a wide range of tasks. It's the failure modes above that determine when it's the right tool and when it isn't. We're not saying "AI is overhyped" as a broad claim. Some of the hype IS overhype; some is real. The right response is calibrated use, not blanket cynicism. We're not saying "wait for GPT-5 or Claude 5." The successor models will be better on some of these axes and still fail on others (probably subtler ones). The list gets shorter over time; it never becomes empty. **Why we're publishing this:** Because a product's honesty about its limits is a form of respect for its users. If we didn't tell you, someone would eventually publish the list for us, and the reveal would be worse than the transparency. Use LADLE for what it's good at. Verify what needs verification. Get humans for what humans do better. That's the real user manual. --- ## LADLE vs Claude.ai — for writers specifically. - Slug: ladle-vs-claude-ai-for-writers - URL: https://ladle.chat/blog/ladle-vs-claude-ai-for-writers - Date: 2026-06-05 - Category: NOTE - Summary: Same model. Same price. For writers, the differences that matter are voice-preservation defaults, project structure for long-form work, and where the money goes. If you're a writer choosing between LADLE and Claude.ai directly, the model is identical (both are Claude Sonnet), the price is identical ($20/mo), and the underlying capability on prose is identical. The differences that actually matter for writing work are elsewhere. **Voice preservation.** LADLE's default response style leans thorough; if you're doing draft-and-edit work you'll want to switch it to "Concise" or write a custom instruction that says "preserve my voice; don't smooth out sentence rhythm." Same for Claude.ai — its default is chatty. Both are configurable; both need a minute to tune. The end state feels the same. **Long-form editing.** Both support the 200K context window, so pasting whole manuscripts works on both. LADLE's Projects surface (a container for related chats with shared instructions and knowledge files) maps well to book-level work — put your style guide, character list, and outline as project knowledge, keep the ongoing chapters as separate chats. Claude.ai's Projects work similarly. This is a wash. **The artifact panel.** When you ask for a full rewrite or long output, LADLE's artifact panel lifts the output into a side panel, versioned, so you can iterate across many turns without losing prior drafts in the scrollback. Claude.ai has a similar feature. Also a wash. **What's actually different.** LADLE's monthly $20 sends $8 to the World Food Programme with a public receipt. Claude.ai's $20 stays entirely in Anthropic's economy. If that difference matters to you, LADLE. If it doesn't, Claude.ai is a perfectly good product and I'd rather you use it than reluctantly pay us. **One writer-specific edge case:** if you're editing manuscripts with sensitive personal information (memoirs, unpublished fiction with real people), both products' data handling is fine for the API tier we use (no training on your content). If you need "not even the model host sees it," neither is the tool — running an open-weights model locally is. **Recommendation.** Try both for a week each. The chat interface is more different than the underlying model. Pick whichever workflow you settle into. --- ## Extended thinking, in plain English. - Slug: extended-thinking-in-plain-english - URL: https://ladle.chat/blog/extended-thinking-in-plain-english - Date: 2026-06-04 - Category: NOTE - Summary: Extended thinking lets Sonnet reason for longer before it starts writing the reply. It helps on hard problems and hurts on easy ones. Here's how to know which is which. Anthropic added extended thinking to Sonnet in early 2025. It's a mode where the model generates hidden "thinking" tokens before it writes the visible reply. Think of it as the model working through the problem on scratch paper before drafting the answer. LADLE exposes this as a toggle in the composer. The default is off. Here's the honest guide to using it. **What extended thinking does.** When it's on, Sonnet allocates a budget of tokens (LADLE budgets 8,000) to internal reasoning before generating the reply. The reasoning is captured and shown in a collapsed "Thinking…" block above the reply — you can expand it to see how the model got there. The reply itself is often better on hard problems: fewer arithmetic mistakes, tighter logic, better handling of multi-step questions. **What it doesn't do.** It doesn't make the model smarter. It gives an already-capable model more room to check its own work. On easy questions, the extra thinking is wasted (and costs you tokens against your budget). On medium-hard questions, the reply tends to be a hair more careful, sometimes noticeably so. On genuinely hard questions — multi-step math, complex logic puzzles, chained reasoning problems — it makes a real difference. **When to turn it on:** - Math or logic problems where the answer is verifiable and the intermediate steps matter - Multi-step reasoning where the model needs to keep track of state across the argument - Complex code where getting the algorithm right in one shot matters more than iterating - Any task where you'd naturally say "take your time to get this right" **When to leave it off:** - Drafting, editing, translation — these don't need extra reasoning, they need voice - Short factual questions — the reply is done before extended thinking would even finish - Conversational back-and-forth — the extra latency (thinking tokens generate before the visible reply starts) makes the chat feel slow - Anything where the answer is stylistic rather than logical **On latency.** Extended thinking makes replies take longer. Not a lot longer — usually 5-15 seconds extra on a typical turn — but noticeable. If you're in a flow of quick turns, that friction adds up. Turn it on for the hard one; leave it off for the flow. **On cost.** Thinking tokens are billed at the same rate as output tokens (Sonnet: $15/million). An 8,000-token thinking budget adds about $0.12 of compute per turn. Not huge on its own but real if you leave it on across many turns. LADLE's metering counts thinking against your monthly budget the same as output tokens. **On seeing the reasoning.** The "Thinking…" block is expandable. Some users find it fascinating; others find it noise. It's collapsed by default. What's worth knowing: the reasoning shown is a compressed representation of the model's process, not a raw transcript — Anthropic surfaces enough for it to be useful without cluttering the reply. **Rule of thumb.** If you'd naturally say "take your time," turn it on. If you want the reply now, leave it off. The toggle is per chat, so a heavy-reasoning research chat can have it on while your daily drafting chat has it off. --- ## Three habits from founders who use LADLE well. - Slug: founders-three-habits - URL: https://ladle.chat/blog/founders-three-habits - Date: 2026-06-01 - Category: NOTE - Summary: Not the productivity-porn version. Three specific daily patterns from early-stage founders that keep the model useful without turning it into a distraction. We've watched enough early-stage founders use LADLE to notice the ones who get real leverage from it and the ones who bounce off. The productive users have converged on three specific habits. This post is those three, described plainly enough to steal. **Habit one: the morning triage prompt.** Every morning, before opening the calendar, the founder pastes their top-of-mind list (three to five things bugging them) into LADLE with the same prompt: "Here's what's on my mind today. Which of these are decisions I should make in the next hour, which are things I should defer to a scheduled block later this week, and which are worries I should just write down and not act on? Be direct." The reply is a triage, not advice. What matters is the reframe: seeing your own list categorized by urgency and actionability, from an outside perspective, in 30 seconds. It kills the "everything feels equally important" morning that eats an hour of productive thinking. The founders who do this say the same thing: the model isn't smarter than they are, but the act of writing the list out and reading it back with a distanced perspective is worth more than the model's specific categorization. **Habit two: pre-meeting prep in three prompts.** Fifteen minutes before an important meeting (investor call, hiring conversation, difficult team discussion), three quick prompts: Prompt 1: "I have a meeting with [person] about [topic]. Given what you know about [context], what are three questions they're likely to ask?" Prompt 2: "For each of those, what would be a bad answer that founders often give, and what's a version that's honest and specific?" Prompt 3: "What's one question I should ask them that I'm probably not planning to?" The output is a two-minute read, done in the elevator or the two-minute walk between rooms. Not a script. A checklist against your own blind spots. Founders who do this walk into meetings with less anxiety and better questions. **Habit three: the end-of-day forcing function.** Last thing before closing the laptop: "Summarize what I built or decided today from these three chat threads: [paste 3 chat titles or IDs]. What did I miss? What's one loose end I should close before tomorrow?" Twofold effect. First, it forces a summary that goes into your own memory (writing it out helps you remember). Second, the "one loose end" question catches the small things that would otherwise silently accumulate into a Monday morning tangle. This one is more effective for founders in high-context weeks — launches, fundraises, hard product decisions — than for slow weeks. Skip on slow weeks. **What NOT to do:** Don't use LADLE for status meetings you'd normally have with a co-founder. The model isn't a co-founder. It won't push back on your BS at the right level. Human accountability is different from AI reflection. Don't ask it to make a decision for you. Ask it to help you see the decision. The founders who ask "what should I do about X" get generic advice. The founders who ask "here's the decision I'm circling, here are the options I see, what am I missing" get useful reflection. Don't paste sensitive investor emails or confidential offer letters unless you've checked LADLE's data controls and are comfortable. Standard rule for any AI tool. Temporary chats are one option; local models are another; a private notebook is a third. **Rule of thumb.** The founders who get the most from LADLE use it as a low-stakes thinking partner three or four times a day, not as an oracle they consult for The Answer. Frequent, small, structured. Not rare, big, generic. Try one of the three habits for a week. If it's useful, keep it. If not, drop it. This is a menu, not a system. --- ## The artifact-first workflow. - Slug: the-artifact-first-workflow - URL: https://ladle.chat/blog/the-artifact-first-workflow - Date: 2026-05-30 - Category: NOTE - Summary: Most chat products treat every reply the same. LADLE lifts the durable outputs — HTML, code files, long documents — into a separate panel so you can iterate on them without losing the chat context. Most AI chat products treat every reply the same: text scrolling in a transcript. That works fine for conversation. It works badly for the moments when the reply IS a thing — a full HTML file, a code module, a long document, a design draft. In a scrolling transcript, "the thing" gets buried under follow-up questions and lost when you scroll. LADLE's artifact system lifts these outputs into a side panel. This post is the workflow. **How it works:** The server-side detection notices when a reply meets the artifact bar: a complete HTML document, a code block over a certain length, a markdown document with structure. When it triggers, the artifact opens in a panel on the right, versioned, with the chat continuing on the left. Every subsequent turn that modifies the artifact creates a new version. The panel has a version selector — click backward to any prior version, click forward to bring back the latest. Nothing is lost. Every version is preserved. The artifact is downloadable (as its native file format), copyable (raw content), and, for design artifacts, previewable at desktop/tablet/phone widths. **What this unlocks:** **Iterative design without losing context.** You draft a landing page. The reply is a full HTML file — it opens in the panel, renders live in the iframe. You say "make the hero smaller and add a testimonial section." Reply lands, new version, panel updates. The chat log stays clean; the artifact evolves. **Code review that produces artifacts.** You paste a file, ask for a rewrite. The rewrite appears as an artifact you can copy in one click. Your notes to the model live in the chat; the actual output is in the panel, ready to paste back into your editor. **Document drafting.** Long-form writing — a policy doc, a spec, a proposal — lives as a markdown artifact. Edits happen through the chat ("expand section 3 to include the exceptions"), the artifact regenerates, you keep working. When it's done, download as MD. **Reference back to prior work.** Every artifact from every chat is browsable at /artifacts. Filter by type, search by title, jump back to any prior version. The artifact library is a growing archive of the durable outputs you've created with LADLE. **Why the panel design:** We considered inline (artifacts embedded in the chat transcript, like ChatGPT canvases). Two problems: One, they get lost when you scroll. Ten turns down the transcript, the artifact from turn one is behind ten pages of chat. Two, they compete with the chat for visual attention. If the artifact is prominent inline, it's fighting the message that produced it. If it's small inline, it's not useful. The panel model — dedicated space, on-screen persistently, versioned — solves both. Left side is the conversation; right side is the output. Both visible; both scrollable independently. **What's an artifact and what isn't:** Artifacts trigger for: - Complete HTML documents (single-file, self-contained) - Code blocks over ~40 lines with clear file structure - Long markdown documents with headers and structure - Design outputs (specifically flagged by the design system prompt in design mode) Artifacts do NOT trigger for: - Short code snippets (a 5-line function stays inline) - Explanatory text with a code example in it - Casual replies - Follow-up questions The bar is: "is this a thing the user would want to iterate on separately, or is this part of a conversation." The former is an artifact; the latter is a message. **One caveat:** The artifact detection isn't 100% accurate. Sometimes it fires when it shouldn't (a code sample gets promoted to an artifact when a normal message would have been fine). Sometimes it misses (a long doc that would benefit from being an artifact stays inline). The heuristic is under active revision. If you get a stale detection, closing the panel doesn't lose the reply — it's still in the transcript. The rule: artifact = the thing you'd copy, iterate on, or ship. Everything else is a message. --- ## The kitchen panel: why every user sees the router. - Slug: the-kitchen-panel - URL: https://ladle.chat/blog/the-kitchen-panel - Date: 2026-05-27 - Category: NOTE - Summary: Most consumer AI products hide the model-selection logic. LADLE ships it as a keyboard shortcut. Here's the reasoning — and what you can actually learn from opening it. Press Cmd+K on any LADLE reply and a panel slides in from the right. It shows: which model produced the reply, why the router chose that model, how many input and output tokens the turn consumed, whether search or extended thinking fired, and a rough per-turn cost estimate. Every LADLE user has access to this. It's not a power-user setting or a paid tier. It's a keyboard shortcut on every reply. **Why we built it.** The AI industry has a transparency problem. When you ask ChatGPT a question, you don't know which model actually answered — GPT-4, 4o, 4o-mini, the routing rules aren't published. When Claude.ai routes for you, same story. The user is opting into a system that decides on their behalf, without seeing the decision. That's fine for most users most of the time — the routing is generally reasonable and the reply arrives. But it becomes annoying when the reply feels off ("did I get the small model?"), when the bill seems high ("what did I use these tokens on?"), or when you're trying to understand the product you're paying for ("what does 'premium' actually mean here?"). The kitchen panel is our answer. It shows the routing decision, the token counts, and the reasoning. If you got a Haiku reply on a hard question and the reply was disappointing, the panel says "routed to Haiku because the prompt looked short and factual." If you're wondering why extended thinking added 12 seconds of latency, the panel says "thinking budget 8,000 tokens, actual thinking output ~5,200 tokens." **What you can learn from it.** Some patterns users have discovered by watching the kitchen panel: Rewriting a short prompt as a longer, more detailed one often shifts the router from Haiku to Sonnet. The router uses prompt length and complexity as one signal. If your quick question kept getting a shallow reply, the kitchen panel tells you why. Search-on turns cost more, and the panel shows how much. If your budget is tight this month and you notice most of your turns show 1-3 searches, you can turn search off deliberately for turns that don't need it. Extended thinking rarely fires on drafting tasks, even when you leave it on — the model chooses whether to actually use the budget. The kitchen shows how many thinking tokens got spent, which is often zero even with thinking enabled. **Why not hide it behind a "developer mode" flag.** Because it isn't a developer feature. It's information the user paid for. Hiding it behind a toggle would communicate "this is complicated and probably confusing" — which we don't think is true. The panel uses plain English. Model names are model names. Tokens are labeled. The reasoning is one sentence per decision. The one thing the kitchen panel isn't: a way to override the router. If you want a specific model, use the model chip on the composer (Fast forces Haiku, Deep forces Sonnet). The panel is for seeing what happened, not for controlling it — for that we have explicit controls, one click away. Software should let you look under the hood. LADLE just makes the hood open by default. --- ## Migrating from ChatGPT: what to expect. - Slug: migrating-from-chatgpt - URL: https://ladle.chat/blog/migrating-from-chatgpt - Date: 2026-05-23 - Category: NOTE - Summary: The move from ChatGPT to LADLE is smaller than you'd think. Here's the honest guide — what stays the same, what feels different, and the two things that'll frustrate you in week one. A meaningful fraction of LADLE subscribers came from ChatGPT. This post is the honest migration guide, informed by what those users have told us about the transition. **What'll feel similar:** The chat surface. It's a chat with an AI assistant. You type, it replies, you type again. You can attach files, ask for code, get long-form writing. The core interaction pattern is basically the same across every major consumer AI product now, and LADLE isn't trying to reinvent it. The model quality on everyday tasks. Claude Sonnet and GPT-4 (or GPT-5, depending on which OpenAI tier you were on) are close enough on typical assistant work that most users describe the switch as "different feel, similar competence." Claude has a slightly more literary voice; GPT is a bit more matter-of-fact. Preference thing, not a quality thing. The file support. PDFs, images, code, documents — all handled. Attachment limits are comparable. **What'll feel different:** The voice. Claude tends toward more thorough replies, occasionally more cautious framing, slightly more literary phrasing. If you loved ChatGPT's tighter default replies, Claude may feel a bit chatty at first. LADLE's response-style settings let you tune this (concise / thorough / exploratory / precise) — set it to "concise" in Settings if you want shorter replies out of the gate. The routing transparency. LADLE tells you which model produced every reply and why (kitchen panel, Cmd+K). ChatGPT hides its routing. For most people this doesn't matter; for some it's a delight. The composer. LADLE's composer has explicit toggles for search, thinking, temporary chat, and model tier. ChatGPT has some of these but bundled differently. You'll relearn the muscle memory in a day. The meals. Every subscription earmarks $8 to WFP each month. This isn't a feature you interact with day-to-day, but it's the reason LADLE exists as a separate product from Claude.ai (which is same price, same model, no meals attached). **Two rough patches to expect in week one:** **Rough patch one: the reply length.** Claude Sonnet's default replies are meaningfully longer than ChatGPT's defaults for the same question. This is a model-family thing, not a LADLE thing. Turn on the "concise" response style in Settings → Behavior for the first week to calibrate. After a week, most users find the more thorough default helpful; a few keep concise forever. **Rough patch two: the projects behavior.** If you were using ChatGPT Projects heavily, LADLE has projects too but the mental model is slightly different (project instructions apply to every chat in the project; knowledge files are attached at the project level rather than per-chat). Read the docs on projects before you set up your first one — it'll save you two revisions. **Things to export from ChatGPT before you switch:** - Your ChatGPT chat history (Settings → Data controls → Export). You can't import it into LADLE, but you'll want it for reference. LADLE does have full chat export the same way. - Any custom GPTs you built or used. LADLE doesn't have a GPTs equivalent yet; projects are the closest analog but they're not the same primitive. - Any Memory items ChatGPT was holding for you. LADLE has memory too, and the "onboarding" step 6 asks about memory preferences — you can seed your own memory items in Settings → Memory. **What to do the first week:** Set your response style. Set your theme. Try a few of your normal daily chats and pay attention to the kitchen panel — see how the router decides. Do one design chat if you're curious. Poke at Fast vs Deep. That's about it. **One thing to NOT do:** Don't cancel your ChatGPT subscription on day one. Let them coexist for 30 days. Some users find they prefer LADLE for most things and keep ChatGPT for a specific workflow (usually image generation, which LADLE doesn't have, or a specific custom GPT they built). That's a legitimate outcome — the two products aren't perfectly overlapping. Most people who stick with LADLE beyond 30 days end up canceling ChatGPT within another month. But you don't have to commit on day one, and the meal donation from your LADLE subscription is happening regardless. --- ## Why we won't chase seats. - Slug: no-seat-pricing - URL: https://ladle.chat/blog/no-seat-pricing - Date: 2026-05-19 - Category: NOTE - Summary: ChatGPT Business charges $25/seat/month. Claude Team charges $30. LADLE charges $20 whether you're a solo user or seat 42 of a company subscription. Here's the specific reasoning. The consumer AI vendors all have a business tier now, and they all price it higher per seat than the consumer product. ChatGPT Business is $25/user/month on an annual commitment. Claude Team is $30 monthly or $25 annual. Copilot for Business is $19/user. The convention is: consumer product minus a few features, plus admin controls, times more dollars per person. LADLE for business is the exact same product at the exact same price. $20/user/month, no annual commitment, no minimum seat count, no per-seat markup. Buy one seat or five hundred; same price per user. We give up money doing this. If we'd priced business at $25/seat we'd make an extra $5 × business subscriber every month. At even 5% business mix at scale, that's real revenue we're leaving on the table. Two reasons we do it anyway. First, we already earmark $8 of every $20 for WFP. Charging $25 to businesses would either raise the meal count for their subscribers (creating a two-tier meal donation that fractures the marketing) or increase our margin (which we already say is fixed at ~$3.50/subscriber). Neither is honest. The uniform price protects the uniform meal promise. Second, the industry price ladder for business AI reflects features consumer plans don't get — admin dashboards, audit logs, SSO, data controls, procurement contracts. We ship those to everyone. Every LADLE subscriber gets the same data controls, the same export tools, the same account security. There's no locked door for the business tier to unlock. Charging more for the same product would be dishonest markup. What businesses do get that solo users don't: an impact report on demand (PDF, CSV, or JSON, dated and referenced) that the CFO or CSR team can hand to a stakeholder. That report is a real deliverable, not a feature we invented for tier justification. It costs us $0 to generate — it reads from the same ledger the receipts page reads from — so charging for it would be strange. We also don't do a minimum seat count. Some vendors require 2 or 5 seats for business features to unlock. Doesn't apply here — a two-person company gets the same product and the same impact report as a five-hundred-person one, with the same $20/seat price and the same $8/seat/month to WFP. The loss here is real: we're smaller than we could be with a business-tier upcharge. The gain is that the deal is clean. "LADLE for business" isn't a bigger version of LADLE. It's LADLE, ordered on a company card, with a report on request. If we ever do add a business-only feature that costs meaningful money to run (a dedicated Slack integration, single-sign-on with a niche IDP, a compliance certification), we'll price it as an add-on, not as a per-seat multiplier. The base subscription stays $20. --- ## Aggregating micro-donations: why the batch matters. - Slug: aggregating-micro-donations - URL: https://ladle.chat/blog/aggregating-micro-donations - Date: 2026-05-14 - Category: NOTE - Summary: Ten thousand $8 wires per month would waste 15% in fees and swamp WFP's operations. One monthly batch of the total is more efficient. Here's the math and the timing. Every LADLE subscription earmarks $8 for the World Food Programme. If we naively sent each $8 as an individual wire transfer as the Stripe charge cleared, three things would happen: (1) we'd burn a huge fraction of the donation in transfer fees, (2) WFP's finance operations would drown in tiny incoming transactions, (3) the receipts would be per-transaction chaos rather than per-month clarity. So we batch. Every month, on the last day, we send one aggregated wire transfer to WFP through ShareTheMeal. This post is the specific reasoning about the batch, since it comes up occasionally. **The transfer-cost math.** International wire transfers have fixed costs — typically $15-50 per transaction depending on the corridor and bank. Even if we used a low-cost cross-border rail (Wise, or similar) fees are meaningful on small transactions. At $8 individual transfers, a $10 fee is 125% of the donation being consumed by overhead. At a $500 aggregated transfer, the same $10 fee is 2%. At a $10,000 aggregated transfer, it's 0.1%. Batching is not optional at scale. **The operational math.** WFP and ShareTheMeal are not designed to receive thousands of individual transfers per corporate donor per month. Their systems can, but their finance staff shouldn't have to. When we aggregate, they see one incoming transfer with a summary line item, cross-reference it to our monthly report, issue one receipt on their end, and move on to distribution planning. Individual per-subscription transfers would create reconciliation overhead on their side that would consume WFP staff time — which is the opposite of what we want. Their staff time should go to logistics, not to paperwork about our micro-donations. **The timing.** Charge clears (say, on the 3rd of the month) → $8 earmarked immediately in the LADLE ledger to "pending accrual for month X" → at end of month, sum all accruals for the month, subtract any refunds that occurred in the 7-day window, wire the net total to WFP → receipt published within 5 business days after the wire clears. This means: your July 3rd subscription's meal donation is part of the July batch, wired at end of July, receipted in early August. Timing lag from your charge to your meals arriving in WFP's operational budget: 3-4 weeks. **What about refunds mid-month?** Refunds within the 7-day window unwind the meal accrual before the batch is sent. Refunds after the batch has been sent don't get reversed on WFP's side — the meals were funded — and LADLE absorbs the $8 as an ops cost. This has happened rarely enough that the operational cleanliness (never asking WFP to unwind a batch) is worth the small absorbed cost. **Why not weekly batches?** We considered. Monthly matches WFP's own operational cadence for how they think about incoming aggregated donor flows. Weekly would create 4x the accounting artifacts for negligible speed improvement. The meals aren't served the day the wire clears anyway — WFP's distribution runs on its own schedule. **The escape hatch.** If the monthly batch doesn't send for any reason (bank issue, our error, WFP receiving delay), we say so in the changelog with the reason. The accrual doesn't disappear — it rolls to the next month with a note explaining the delay. Never happened to date. Documented in advance because it might, and we want to be ready. **Bottom line.** Batching is boring plumbing. It's also the operational move that makes the whole model work at scale. Without batching, either the transfer fees would eat 15%+ of every donation, or WFP would be drowning in reconciliation work — either way, less meals for the same money. Batching is boring, batching is honest, batching is how the numbers stay big. --- ## AI for writers who don't want a robot voice. - Slug: writers-who-dont-want-a-robot-voice - URL: https://ladle.chat/blog/writers-who-dont-want-a-robot-voice - Date: 2026-05-08 - Category: NOTE - Summary: The AI-assisted writing you can spot from a mile away is written by people using AI as a first-drafter. There's another way to use it — as a second pair of eyes — that doesn't cost your voice. Every writer I know has tried using AI for drafting. Most of them stopped. The reason is usually the same: whatever came out of the model sounded like the model, not like them, and the effort to un-model-ify it was more than the effort to just write the thing. There's another way to use AI for writing that doesn't have this problem. Three specific moves. This is the shape of it. **Move one: use it as a second pair of eyes on your draft, not as a first drafter.** Write the piece yourself. Not a rough outline for the model to fill in — the actual draft, in your voice, at whatever length you'd normally write. Then paste it and ask the model something specific: "Where does the argument thin out?" "What did I probably assume the reader already knows?" "What's one paragraph I could cut?" The reply is a note, not a rewrite. You do the rewriting. The model's job is to notice things you're too close to notice. This is how a good editor works. The model isn't a good editor at that level, but it's a competent one, and it costs $0.05 per pass instead of $500. **Move two: use it for the structural work you'd otherwise skip.** Some kinds of writing benefit from structural attention you don't naturally give them. Cold emails. Job cover letters. Grant applications. Product pitches. There's a "structure of a good X" for each of these, and the model knows it. Give the model the draft and the goal ("this is a cover letter for a senior editor role"), and ask it to check your draft against the structural expectations of the form. The output is diagnostic ("your third paragraph does the work your second paragraph should have done") not rewriting. You then rewrite in your voice with the structural note in mind. **Move three: use it for the parts where the voice doesn't matter.** Some parts of your writing genuinely don't need to sound like you. Metadata. SEO subtitles. Email subject lines. Alt text. Cross-referenced citations. Formatted lists at the bottom of a longer piece. For these, the model's neutral competence is a feature, not a bug. You'd normally write these last, tired, in five minutes. The model does them in ten seconds. Voice unaffected. **What NOT to do:** Don't ask the model to "make this sound more like me." It can't — it has no baseline for what you sound like beyond what's in the current context, and its notion of "your voice" is going to be a caricature of a few style markers rather than the whole texture. If you want more of your voice, cut lines and rewrite them yourself. Don't ask for "a punchier version" or "a tighter version" of a draft you like. You'll get a shorter, punchier, blander version. The model's compression bias is toward safe, common phrasing. Your specific choices are the thing that makes writing yours; compression removes them. Don't paste the model's rewrite back into your document. If you liked something the model did, retype it in your own words. Sounds silly. Isn't. **Rule of thumb.** The rule of thumb: the model is a tool for making your writing better in the places you'd already improve if you had unlimited attention. It's not a shortcut past the writing work. Writers who use it that way keep their voice. Writers who use it as a first-drafter lose theirs. The best test: read three published pieces from someone you know uses AI. Can you tell which ones? If the answer is yes on all three, they're on the wrong side of this. If the answer is "one, maybe" or "no," they're doing it right. --- ## Why we don't have a free tier. - Slug: why-no-free-tier - URL: https://ladle.chat/blog/why-no-free-tier - Date: 2026-05-06 - Category: NOTE - Summary: Free tiers are the industry default for a reason — and they break when the point of the paid tier is that some of the money leaves. Here's why we said no. The consumer AI industry runs on free tiers. ChatGPT has one, Claude.ai has one, Gemini has one, Perplexity has one, DeepSeek is entirely free. It's the default, and it's the default for a reason: free users convert at a low but real rate to paid, and the free tier acts as a marketing funnel that costs less than paid acquisition would. We got asked about a free tier four times in the first month. The math didn't work then and it still doesn't. Here's the specific reasoning. The point of LADLE is that $8 of every subscription leaves the account each month for a specific WFP donation with a receipt. This is the whole product-market fit — not the chat interface, which is fine but unremarkable; not Claude Sonnet, which anyone can access; not the writing quality, which is a wash against competitors. The meals are the reason to pay us instead of the alternatives. A free tier would need one of two things. Either the free users would also fund meals (which would mean we pay Anthropic $6-$7 in API cost per free user's typical monthly usage AND pay $8 to WFP on their behalf — that's -$14/user/month, unsustainable). Or free users would not fund meals, which fractures the pitch: LADLE becomes "the AI subscription where sometimes some users fund some meals" instead of "the AI subscription where subscribers fund ten meals a month, minimum, with a receipt". The second version is a marketing story with an asterisk. Marketing stories with asterisks lose their power. We chose to keep the clean version. There's a second-order effect too. Free tiers create pressure to move features to paid to justify the price, which puts you on the treadmill of paid-tier feature engineering rather than product refinement. We'd rather have one product that does one thing and be good at that. The demo (three free messages, no signup) is our compromise. It lets you try LADLE without committing, and it doesn't require us to sustain a free user's ongoing usage. If the three-message demo isn't enough to decide, that's usually a signal LADLE isn't right for you — and that's fine, ChatGPT's free tier is a real product and we're happy to recommend it. The strongest argument for a free tier is that it dramatically increases the subscriber count at scale, which increases the total WFP donation. That's a real argument. We think a smaller number of committed, paying subscribers who understand the deal produces a healthier company than a large free base with a conversion-optimization treadmill. Ten years from now, this decision either looks prescient or looks like we left significant impact on the table. We're comfortable being judged either way. --- ## Web search: when to leave it on. - Slug: web-search-when-to-leave-on - URL: https://ladle.chat/blog/web-search-when-to-leave-on - Date: 2026-05-02 - Category: NOTE - Summary: Web search is a toggle in the composer, not a default. Here's the practical guide to when it helps (fresh facts, cited sources) and when leaving it on actively hurts your reply. LADLE has a web search toggle in the composer. It's off by default. Deliberately. Web search is expensive — Anthropic charges a small fee per search, and a heavy-search subscriber can burn through their compute budget faster than a same-usage subscriber who leaves search off. That's the direct cost. There's also an indirect one: web search changes how the model behaves in ways that aren't always what you want. When you leave search ON, the model treats every question as if it might need to look something up. That's great for "what's the latest on the CPI print" and terrible for "help me write a birthday message for my sister." In the second case, the model will do a search, find nothing useful, and produce a subtly stiffer reply than if it had just answered from what it knows. Here's when to turn search ON: - Facts that change over time: current events, prices, availability, versions - Cited references you can click: research assistance where you want URLs, not just claims - Company or product research: model's knowledge cuts off; recent moves won't be in training - Anything where "the model might hallucinate a URL" is a live risk Here's when to leave search OFF: - Writing drafts, edits, translations, tone work - Code that doesn't depend on a specific library version - Explaining concepts, working through problems, exploratory thinking - Personal or subjective tasks (advice, planning, brainstorming) The rule of thumb: if the answer would be different in six months than it is today, turn search on. If the answer is timeless, leave it off. There's a settings-level default too. Some users work in domains where search-on is right most of the time (journalists, researchers, competitive analysts). You can set search-on as your default in Settings → Behavior. It'll still be a toggle per chat if you want to flip it. One caveat about how search works in LADLE. The tool is Anthropic's native web search — the model itself decides whether to invoke it, and how many queries to run per turn (we cap it at three per turn so a runaway search chain doesn't eat someone's budget). When it searches, you'll see a "SEARCHING · query" line render in the reply, then the answer with citation footnotes at the bottom. Citations are real URLs from the search results, not fabricated ones. If you ever see a citation URL that 404s, that's Anthropic's search index temporarily out of sync with the live web — report it and we'll pass it along. The composer toggle is one click. Get in the habit of flipping it on for research turns and off for drafting turns. Two months in, the users who report best experience are the ones who use search deliberately, not the ones who leave it on or off permanently. --- ## Why we're a public benefit corporation. - Slug: why-were-a-pbc - URL: https://ladle.chat/blog/why-were-a-pbc - Date: 2026-04-30 - Category: NOTE - Summary: LADLE PBC is a Delaware public benefit corporation. That's a real structural choice with real implications. Here's what it means, and what it doesn't. LADLE is incorporated as LADLE PBC, a Delaware public benefit corporation. The "PBC" isn't a marketing tag — it's the legal structure. This post is what that actually means. **What a PBC is:** A public benefit corporation is a for-profit company whose charter commits it to a specific public benefit purpose alongside profit. In Delaware (where LADLE is incorporated, like most US startups), a PBC has all the powers of a normal C-corp — can raise capital, issue stock, be acquired — but its board is legally bound to consider the stated public benefit purpose when making decisions, not just shareholder value. Our stated public benefit is roughly: contributing to global food security by directing a fixed portion of revenue to nutritional assistance programs operating at scale. **What a PBC is NOT:** It's not a nonprofit. LADLE has shareholders, can distribute profits, can be sold, can raise venture capital (though we haven't). Nonprofit status is a different structure entirely (501(c)(3) in the US), with different tax treatment and different governance constraints. PBCs pay taxes normally. They're not tax-deductible to donors because they aren't charities. It's not a certification. B Corp is a certification (issued by B Lab) that any company structure can pursue. PBC is a legal structure. Some B Corps are PBCs; many aren't. Some PBCs are B Corps; many aren't. LADLE is a PBC but hasn't pursued B Corp certification (yet or maybe ever — B Lab's process is meaningful work and we haven't prioritized it). It's not marketing polish. The PBC filing is a public document (Delaware has a searchable registry). The obligation is real, enforceable in Delaware Chancery Court by shareholders. If LADLE's board approved a decision that materially undermined the public benefit purpose, shareholders could sue. **Why we chose PBC over regular C-corp:** Because we wanted the mission commitment to be structural, not just cultural. In a normal C-corp, a hostile board or an acquiring company could unwind LADLE's meal donation commitment tomorrow and every shareholder would have to accept it. In a PBC, doing that would breach the charter, and the board would face legal liability. The commitment is protected from future drift. There's also a signaling reason. When a company says "we care about impact," you can only judge by their actions. When a company files as a PBC with a specific public benefit purpose in the charter, you can go read the charter and judge from the document. Structural claims are more credible than cultural claims. **Why we chose PBC over nonprofit:** Because we want to be a normal software company with a specific unusual policy, not a charity that runs software. The difference matters: - Nonprofits have to run on donations and grants, with all the operational overhead that comes with fundraising, board development, and 501(c)(3) reporting. - Nonprofits can't easily raise venture capital, which limits the paths to scale. - Nonprofit governance is designed for stewardship of charitable assets, not for running a competitive consumer software product. We wanted to compete with ChatGPT and Claude.ai on product terms while doing the specific thing with money we do. That's a PBC's job description; not a nonprofit's. **What being a PBC changes about how LADLE runs:** The board's fiduciary duty explicitly includes the public benefit purpose. If we ever raise money, that constraint doesn't go away — investors buy into a PBC knowing the public benefit commitment is durable, not something they can vote to unwind for a bigger exit. The annual report to shareholders has to include a public benefit statement — what we did, how we measured it. The receipts page is essentially the operational form of this reporting. If we ever considered acquiring or being acquired, the transaction structure has to respect the public benefit purpose. We can't be silently absorbed into a company that then quietly drops the meals. **What it doesn't change:** Nothing about the day-to-day experience of using LADLE. The product is the product. The PBC structure is scaffolding that makes the promise durable across leadership changes, ownership changes, and time. If you want to read the actual charter, it's at /about/pbc-charter. If you want to verify the filing, it's on the Delaware Division of Corporations' public database under LADLE PBC. --- ## Refunds without questions. - Slug: refunds-without-questions - URL: https://ladle.chat/blog/refunds-without-questions - Date: 2026-04-25 - Category: NOTE - Summary: Seven days. Any reason or no reason. One email. Refunds reverse the meal booking too — WFP handles it through the partner platform. The whole thing runs on the honor system for both sides. We refund any LADLE subscription within 7 days of purchase. Monthly or annual, doesn't matter. One email to hello@ladle.chat and it's done, usually in an hour. No form. No "please tell us why so we can improve." No retention offer. No pop-up trying to talk you out of it. The point of writing this post is to say — explicitly — that the friction most SaaS refunds carry is a deliberate design choice, and we chose the other one. Here's the reasoning. If you subscribed and the product isn't for you, keeping your $20 hostage costs you $20 and costs us goodwill worth much more than $20. The customer who tried LADLE, didn't love it, and got their money back cleanly is the customer who mentions us on a podcast in 2027 when someone asks about AI subscriptions with a social contract. The customer who felt jerked around during the refund isn't. If you subscribed and the product IS for you but this month is a bad month, we don't want the money. Software subscriptions where you're not using the software but feel too awkward to cancel is one of the ugliest patterns in consumer tech. We'd rather have you back later than have you resenting us now. There's a design constraint in this too. A pain-free refund makes it possible for us to say honestly: try LADLE for a week, and if it isn't your assistant, get your $20 back cleanly. That sentence carries the demo's job of "let me see if this works for me" and then some. What happens to the meal donation when you refund? We reverse it. WFP's ShareTheMeal platform handles the return through their partner infrastructure. If you subscribed on the 3rd and refunded on the 6th, the $8 that was earmarked for that month's donation batch unwinds before the batch is sent. No pretend meals accrue on your behalf; no meals get double-counted somewhere else. If you refund AFTER the monthly donation batch has been sent to WFP, the meals were funded — we don't ask WFP to unwind their side, and we absorb the $8 as ops cost. This has happened maybe six times since launch and it's fine. The clarity of the refund is worth more than the $48 across those cases. Two things we do NOT refund: - Beyond 7 days from purchase on a fresh subscription. You get the paid period out. - On annual plans: same 7 days from purchase. After that, you keep access through the year and the meals stay booked. The policy is on the /legal/refunds page in the plain version. This post is the reasoning. If you're reading this because you're on day 5 and unsure, email us. We'll refund without asking why. Most people who refund never write back to explain, and that's fine — we're not owed the explanation. --- ## How our onboarding stayed at 10 steps (not 3, not 30). - Slug: ten-onboarding-steps - URL: https://ladle.chat/blog/ten-onboarding-steps - Date: 2026-04-19 - Category: NOTE - Summary: The industry gospel is 'onboarding must be short.' We kept LADLE's at ten steps and made every step skippable. Here's why longer worked better than shorter for the problem we were solving. Every startup product blog says the same thing: shorten your onboarding. Three steps. Zero if possible. Get users to value before they can bounce. We tried the short version first. A single "sign up and start chatting" flow with everything deferred to settings. Users signed up, started chatting, and about 20% never came back after the first session. The chats they had were fine — normal quality, normal length — but they weren't the chats that made someone stay. The specific thing that was missing: users who signed up hadn't yet told the model who they were or what they wanted from it. So the model produced generic, appropriate replies to whatever they asked, and those replies felt like every other AI chat. Nothing wrong with them. Nothing memorable either. So we tried the long version. Ten steps, called "Setting the Table," walking through: 1. Your name (how the model should address you) 2. Your role (student, developer, writer, founder, etc) 3. Your intents (what you're going to use LADLE for) 4. Your response style (concise, thorough, exploratory, or precise) 5. Theme choice (midnight, daylight, or system) 6. Memory preferences (opt in or out of the model remembering personal facts) 7. Response language (English, or specify your default) 8. Custom instructions (free-form voice guidance) 9. A first-prompt sandbox to try LADLE with everything set 10. Plan intent (which plan sounds right for you — no commitment) Every step is skippable in one click. If you skip everything, you land in the same place you would have with the short version. But the users who go through them get replies tuned to who they are, and those replies are the difference between "this AI is fine" and "this AI is mine." Numbers from the first quarter after we shipped the long version: Retention week-2 went up materially (we don't share exact figures pre-scale, but "materially" here means the kind of shift that would have shown in a normal A/B test). Median chat length went up — users who went through onboarding write longer, more specific prompts because the model has enough context to reply usefully to specific things. Support tickets went down, especially the "how does this work / what does it do" bucket. The onboarding answered those questions in the doing rather than in the reading. **What we cut from earlier drafts of the flow:** A payment step. We refused to put "add your card" in onboarding. LADLE is a paid product; you subscribe when you're ready, not as a condition of trying it. (The demo is on marketing, three messages, no signup.) A tutorial step. "Here's how to attach files, here's how to start a project." Every product designer will tell you tutorials get skipped. Instead the affordances are visible in the composer and the model itself explains them if you ask. A "connect your accounts" step. Even if we wanted integrations, we wouldn't front-load that ask. Ask for permissions when the user needs the thing that requires them. **The design principle:** Onboarding isn't overhead — it's the first chapter of the product. Short onboarding is right when the product is self-explanatory (a to-do app, a note-taking app). Long onboarding is right when the product's value depends on the model knowing something about you. AI assistants fall in the second bucket. Ten steps, each skippable, is the version that treats the user as an adult and the product as one that gets better the more of yourself you bring to it. --- ## The honest economics of a $20 subscription. - Slug: honest-economics - URL: https://ladle.chat/blog/honest-economics - Date: 2026-04-14 - Category: NOTE - Summary: The specific breakdown of how the $20 flows: $8 to WFP, $7 to Anthropic API inference, $5 to ops. Why the split holds, and what would change it. Six weeks in, the most-asked question after "is the meal donation real" is "how does the $20 actually break down". Here's the full answer. $8 goes to WFP. Fixed portion, not a fraction, not conditional. It's earmarked the moment your Stripe charge clears and lives in a separate ledger from operating funds until the end-of-month transfer. Refunds unwind it; failed charges never enter it. Ten meals per subscription per month at ShareTheMeal's $0.80/meal, with a receipt. $7 goes to Anthropic for API inference. This is the largest operating cost. The average LADLE subscriber sends and receives around 800K tokens per month across all their chats. At Claude Sonnet's per-token pricing on our tier, that averages to $6.20-$7.20 depending on the specific mix of input and output tokens and whether context caching kicks in. We budget $7 as the fair average; heavy users are subsidized by light users. $5 goes to ops. This covers Stripe fees (~2.9% + $0.30 per transaction, so ~$0.90 of your $20 in the first month, less on renewals), infrastructure (AWS + a small SaaS stack, ~$0.60 per subscriber), and roughly $3.50 of margin for the human costs of running this — the founder's salary equivalent (extremely modest at current scale), legal/compliance, and reserves. At 15,000 subscribers, the math works if usage stays close to the averages. If Anthropic drops API prices materially — as they have twice since we launched — we pass that through by extending the average subscriber's usage headroom rather than dropping the $20 price. The $8 to WFP stays constant regardless. Two questions this economics raises that we've heard variations of. "Can you charge less?" Not without cutting the meal donation or losing money. The $20 is the equilibrium price that makes the meal math work at our current usage patterns. If we were subsidized by VC funding, we could — for a while — subsidize a lower price. We aren't, so we can't. "Can you charge more and donate more?" We considered a $30 tier with $16 to WFP. Passed on it because it would require a two-tier product (users on the $30 tier would need to see they were funding more, users on $20 would compare and feel like they were doing less). One plan, one price is a load-bearing constraint. The right answer at scale is more subscribers, not more per subscriber. If any of these numbers change materially, we'll update this post and note the change in the changelog. --- ## What $0.80 buys — the WFP per-meal math, unpacked. - Slug: what-eighty-cents-buys - URL: https://ladle.chat/blog/what-eighty-cents-buys - Date: 2026-04-11 - Category: NOTE - Summary: $0.80/meal sounds too cheap. It isn't. Here's what actually goes into the number — food procurement, logistics, overhead — and how WFP arrives at it honestly. The whole LADLE pitch rests on a single number: $0.80 per meal, via the UN World Food Programme's ShareTheMeal partner platform. Every subscription funds at least ten of them per month. The first reaction most people have is: $0.80 is impossibly cheap. A meal in a rich-country restaurant is $15-30. A grocery-store equivalent is $3-5. How does the world's largest humanitarian food organization deliver a meal for eighty cents? Here's the honest answer. **The meal is one nutritionally-complete portion, not a restaurant meal.** The unit WFP costs at $0.80 is a nutritionally-complete meal for one person in a country where their programs operate. That means: protein source (often legumes, sometimes meat or fish depending on region), starch (rice, wheat, maize, sorghum depending on local staples), and often fortified additions to hit micronutrient targets. It's a real meal — designed by WFP nutritionists to meet caloric and nutritional requirements for a person in the specific context they're being served in. It is not a restaurant experience. It's food that keeps a person alive and functioning at scale. **The $0.80 is the fully-loaded cost, not just the food.** The per-meal number includes: - The actual food (procured locally when possible for both cost and economic-development reasons) - Transport and logistics (which in the countries WFP operates in can be more expensive per calorie than the food itself) - WFP program overhead (staff, security, monitoring, distribution infrastructure) - The ShareTheMeal platform's operating costs That last item matters. ShareTheMeal is the WFP-run platform through which LADLE aggregates donations. Their overhead comes out of donations too — they're transparent about this. When we say $8 buys ten meals, it's inclusive of overhead, not "$8 is the food cost pre-overhead." **Why is it cheaper than a supermarket meal?** Two big reasons. Bulk. WFP procures at massive scale. Their annual food purchases are among the largest in the world. When you buy 100,000 metric tons of maize at once, you pay per-ton prices no consumer ever sees. Geography and product mix. The countries WFP serves have lower food costs than rich countries, and the specific foods they distribute (staples like maize meal, rice, lentils, fortified cereals) are cheaper per calorie than the range a Western consumer eats. Both are real and honest reasons for the low unit cost. Neither is a marketing sleight-of-hand. **Why isn't it even cheaper?** Logistics. In countries where WFP works — often conflict zones, disaster areas, or places with weak infrastructure — moving food from ports or procurement centers to distribution points is expensive and dangerous. Cold-chain logistics for perishables, security escorts through unstable regions, staff hazard pay: these are real costs that WFP absorbs so the meal actually reaches a person. There's no path to a $0.05 meal because the logistics don't shrink to that cost. There's also no reason to inflate the price beyond what it actually is. $0.80 is the honest average. **The upshot for LADLE.** Every subscription earmarks $8. That $8 buys ten meals, most months, at WFP's honest per-meal cost. The meals reach people who otherwise wouldn't have them. The receipt is published monthly with a reference number. If the per-meal cost ever shifts materially (either up or down), we update the meal count in the marketing and the promise. As of this post, $0.80 is what it is, and ten meals is what your $8 buys. --- ## Why Haiku isn't the default. - Slug: why-haiku-isnt-default - URL: https://ladle.chat/blog/why-haiku-isnt-default - Date: 2026-04-08 - Category: NOTE - Summary: Haiku is 1/3 the cost of Sonnet and about twice as fast. On paper it's the obvious default. In practice, defaulting to it would make LADLE worse at most of what people actually ask. Anthropic ships two tiers of Claude that LADLE uses: Sonnet (larger, slower, more expensive) and Haiku (smaller, faster, cheaper). By raw economics, defaulting every reply to Haiku would save us money — inference cost would drop from ~$7/subscriber/month to ~$2-3. We don't. Sonnet is the default. Here's why. Haiku is genuinely excellent at what it's built for: short, well-scoped tasks with clear instructions. Draft a subject line. Classify these three sentences. Answer this factual question from the context I gave you. On tasks like that, Haiku is fast enough to feel instant and reliable enough to feel like the right tool. Haiku is not as good at the tasks that make an assistant feel like an assistant. Long-form drafting where you want the model to read tone and mirror it. Multi-step reasoning where the second step depends on getting the first step right. Code review where the interesting problem is what the code SHOULD do, not what it does. Rewrite-in-my-voice work where the model needs to hold a lot of context and produce coherent output. LADLE users are paying $20/month for a daily assistant. When you open the composer and type "help me think through this," you expect the model that thinks better. If we routed to Haiku by default to save inference cost, the product would feel worse for the exact people paying for the exact reason they're paying. So Sonnet is the default. What does Haiku do at LADLE? Two things. **Fast mode.** There's a toggle in the composer to force Fast, which routes to Haiku. Use it when you want speed over depth — a quick lookup, a short reformatting job, a "just answer this question directly" task. The reply comes back in one or two seconds instead of five or ten. **Router downgrades under load.** When a subscriber approaches their monthly compute budget (this is the metering system that lets us keep pricing at $20 without individual users draining a month in a week), the router downgrades to Haiku instead of blocking sends. Deep model paused; Fast still yours. This is the honest version of "graceful degradation" — you keep getting help, just from the smaller model. The router also has a few auto-Fast cases where Haiku is actually the right answer, not just the cheap one: - Titling chats (30 tokens out, deterministic pattern) - Some routine formatting tasks - One-liner factual questions where Sonnet would over-explain The math on this: about 12-15% of turns end up on Haiku in a typical week, either by user choice or auto-routing. That saves us a modest amount of inference cost — not the difference between profitable and not, but enough to keep the numbers honest. If Haiku ever closes the quality gap on assistant tasks — Anthropic ships new versions regularly, and each Haiku release has been notably better — we'll re-evaluate the default. Not there yet. Until then, Sonnet is the answer for the assistant, and Haiku is the answer for the specific jobs it was built for. --- ## Publishing the first receipt. - Slug: publishing-the-first-receipt - URL: https://ladle.chat/blog/publishing-the-first-receipt - Date: 2026-04-05 - Category: NOTE - Summary: First receipts are the moment the promise becomes a document. Here's the specific checklist we ran to get the first one right — and what we improved for the second. March 2026 was LADLE's first full month of subscribers. April 5th was the day we published the first monthly receipt for WFP donations. This is the specific checklist we ran through to get it right, written down partly for our own future selves and partly so anyone else building a public-receipts model has a starting point. **The checklist (in order):** **1. Reconcile Stripe to the ledger.** Every successful Stripe charge in March that wasn't refunded needs to have contributed $8 to the meal accrual. Every refund needs to have subtracted its $8 accrual. We ran the reconciliation query and matched it to Stripe's monthly summary. Off by $32 on the first pass (four refunds we'd missed in the accrual code). Fixed the code, backfilled the four missed reversals, matched. **2. Compute the aggregate.** Total accrued dollars for the month = total_subscribers × meals × $0.80. Sanity-check against the ledger sum. Both computed independently, both should match. They did after step 1. **3. Send the wire to WFP via ShareTheMeal.** Their platform issues a receipt with a reference number. We wire the aggregate amount. The confirmation comes back with the reference number and confirmation of the batch inclusion. This is the number that anchors the whole receipt. **4. Draft the public receipt page.** Fields: month, subscriber count, total meals funded (subscribers × 10), total USD sent, WFP reference number, batch identifier, per-plan breakdown. Publish at /impact/reports/2026-03. **5. Cross-check the numbers one more time.** Read the receipt as if you're a subscriber checking whether the numbers add up. Do the meals × $0.80 math externally. Match the total to the ledger and the Stripe summary. Confirm the reference number is real by looking it up on ShareTheMeal's side. **6. Publish and announce.** Receipt page live. Note in the changelog with a link. Short update in the blog. No fanfare — the receipt IS the announcement. **7. Archive the underlying artifacts.** Stripe report PDF, ledger CSV, WFP confirmation, batch documentation — all saved to a private drive with the month's identifier. If we ever get audited or challenged, this is the evidence trail. **What we caught before publishing:** The four missed refund reversals (step 1). Would have overstated the meal count by 40. Publishing that number wouldn't have been fraudulent — the total transferred to WFP would have been higher than the refund-adjusted count — but the count would have been inconsistent with the ledger and someone would have noticed. A rounding artifact where the per-plan breakdown didn't quite sum to the total (off by $0.03 due to floating-point rounding across many small transactions). Trivial in impact, would have looked sloppy in the receipt. The reference number format we'd originally displayed was truncated. Fixed to show the full ShareTheMeal reference. **What we changed for month two:** The reconciliation script now runs automatically on the 1st of each month, not manually. Human error → script. The refund reversal test is now a self-test that runs against a synthetic batch every night. Catches the class of bug that hit us in month one before it can hit us in month N. The receipt now includes a "how to verify this" section explaining how a subscriber can independently cross-reference the reference number on ShareTheMeal's platform. We should have had this from the start; adding it made the receipt feel more like a receipt and less like a marketing page. **The general lesson:** The first version of anything public-facing is where you learn what to check for. The checklist we ran on month one was longer than any subsequent month; each month gets shorter as the automation catches more of the surface area. That's how the receipt discipline gets sustainable. The alternative — publishing a receipt that turns out to be wrong six months in — would kill the whole product. The first-receipt care is the deposit against future error. --- ## Annual billing math: 2 months free, 10 months' meals. - Slug: annual-billing-math - URL: https://ladle.chat/blog/annual-billing-math - Date: 2026-04-02 - Category: NOTE - Summary: Annual billing at LADLE is $200 for Ladle base ($40 off the sticker). All 100 meals for those 10 months commit to WFP the day the invoice clears. Here's why meals match months paid. The annual plan costs 10× the monthly price. On Ladle base that's $200 instead of $240 over the year — two months off. On Max 5x it's $1,000 instead of $1,200. On Max 20x it's $2,000 instead of $2,400. Same math at every tier. The interesting choice isn't the discount. It's what happens to the meals. **Meals match months paid.** Annual is 10 months' worth of price, so it's 10 months' worth of meals: 100 meals on Ladle base annual (10 × 10), 500 on Max 5x annual (10 × 50), 1,000 on Max 20x annual (10 × 100). All of them commit to WFP in the month you pay the invoice, not spread across the year. If you paid MONTHLY for a full year — twelve separate $20 charges — you'd fund 120 meals over that year. Annual billing gives you two of those months back as savings, and the meal count moves down with the money. This is the honest version: **you pay for 10, meals match**. Two design choices worth explaining. **Why book all 100 meals on the invoice month, not spread them across 12.** The obvious alternative is: accrue 1/12th of the annual meal count each month, matching the monthly cadence. Cleaner accounting on paper, less month-over-month surge. We don't do that. When an annual invoice clears, we book the full 100 (or 500, or 1,000) meals to that invoice month. Two reasons. First, the meal donation follows the money. The $80 (or $400, or $800) goes to WFP as one line in that month's donation batch. Sitting on someone's cash for eleven months in a "future meals" ledger would be a fiction — the money isn't reserved for them, it's already been paid to Anthropic and to ops as we serve their usage across the year. The clean accounting matches reality: money in, meals booked, done. Second, it removes a class of failure mode. If a monthly-drip approach lived in code and something broke in month seven, the last five months of an annual subscriber's meals could silently not accrue. With one-shot booking at invoice time, if it lands, it lands. Nothing to break later. **Why the meals count matches months paid, not months of access.** The alternative is: "you get access for 12 months, we donate meals for 12 months too — the two extra months' meals are a bonus." That'd be a marketing win but a math loss. LADLE's meal donation is derived from the money that came in ($8 of every $20). If we donated 12 months of meals against 10 months of money, we'd be donating $96 for a $200 invoice — that's $16 more meals than the $8-of-$20 formula supports, and the extra has to come out of ops or margin. We chose to keep the formula clean: $8 to WFP for every $20 of price paid. If you pay $200, we donate $80. That's 100 meals. Simple, verifiable, protected against future price/cost drift. **What about cancellation?** If you cancel an annual subscription mid-year, Stripe's default is no refund on the prorated remainder — you keep access through the paid period. We follow that default. The meals we already booked stay booked. The 10 months' worth funded meals; you used part of the 12 months' worth of access. Both sides of the contract were met. The refund policy handles the edge case: cancel within 7 days of purchase (annual or monthly) and we refund the full amount. In that case we reverse the meal booking too — WFP handles the return through their partner platform. Nothing complicated for the subscriber, nothing pretend on the ledger. **Bottom line.** Annual pricing is the plan for people who know they're staying. Two months off, all of that year's meals booked the day you pay. Nothing back-loaded, nothing conditional, no fake "bonus" meals. If you'd rather have the extra two months of meals per year, pay monthly — you'll spend $40 more per year and fund 20 more meals with it. Same rate either way, just a different way to arrange the math. --- ## LADLE vs. Claude.ai — what's actually different. - Slug: ladle-vs-claude-ai - URL: https://ladle.chat/blog/ladle-vs-claude-ai - Date: 2026-03-30 - Category: NOTE - Summary: Both are $20/month, both run on Claude Sonnet, both do the same core thing. The differences are specific and worth being explicit about — starting with the $8 that leaves your account. LADLE runs on Claude Sonnet. Anthropic's own Claude.ai product runs on Claude Sonnet. Both are $20/month for the base plan. The obvious question — asked correctly and honestly — is: what am I paying LADLE for that I couldn't get from Claude.ai directly? Here's the specific answer. **The main difference: $8/month leaves the account.** Of every $20 LADLE subscription, $8 is earmarked for the World Food Programme, paid monthly to ShareTheMeal, receipted publicly. Ten meals per subscription, minimum, every month. This is the whole reason LADLE exists as a separate product. If you don't care about that, Claude.ai is the right choice for you — same model, same $20, but the whole $20 stays in Anthropic's economy. **What Claude.ai has that LADLE doesn't:** - **Direct model access.** You're one hop closer to the model at Claude.ai. In practice this makes zero perceptible difference — LADLE routes to Anthropic's API through their standard SDK — but if you value being on the source product, that's a real preference. - **Claude-first feature timing.** Anthropic ships new features (extended thinking, projects, connections, computer use) on Claude.ai first. LADLE integrates them on our own schedule, which sometimes lags by a few days to a few weeks. As of publication we have projects, extended thinking, artifacts, and search. We don't have computer use. - **Anthropic's proprietary connections and integrations** (Google Drive, Notion, etc.). We haven't wired these; we might, if there's demand. - **The Anthropic developer suite** (API keys, Console, etc.) if you want to also build with the API. That's separate from Claude.ai but it's under the same account. **What LADLE has that Claude.ai doesn't:** - **The meals** (the whole thing). - **A more opinionated chat surface.** Our composer, kitchen panel (which shows the routing/cost per turn), design mode, artifact panel, and onboarding are ours. Not better or worse — different, and opinionated. Some people prefer more spartan; some prefer more considered. - **Full model transparency.** Every reply shows which model produced it, why, and what it cost you. Claude.ai has less of this visible. - **Different theme system.** LADLE has a Daylight mode (warm cream) designed alongside Midnight. Claude.ai has its own dark/light. Preference thing. - **A different data policy on chats.** Both products say chats aren't used for training. Both have the standard trust and safety window. Details are in each product's privacy policy. Read them if it matters to you. **Where they're identical:** - The underlying model. Claude Sonnet's replies are the same replies. No LADLE-specific fine-tuning; Anthropic's model unchanged. - The pricing tier structure at the top. Both go $20 → higher tier with more usage. LADLE's Max 5x is $100; Anthropic's Claude Max is $100 also. Different products but comparable price points. - The 200K context window, the file support, the reasoning capabilities, the safety behavior. **Who should pick LADLE:** - You want a chat product that fits in your normal life AND does something you'd otherwise donate for. LADLE gives you both without adding a step. - You value the specific product opinions we ship (the kitchen panel, the design canvas, the artifact-first workflow for HTML, the honest error copy). - You want to signal that you care about the social contract of the software you use, and the receipt is the artifact that lets you prove it. **Who should pick Claude.ai:** - You don't care about the meal donation and would rather have your $20 go to model development directly. - You want the newest Anthropic features the day they ship, without waiting for our integration. - You want Anthropic's specific integrations (Google Drive, Notion) that we haven't built. **Neither of us is trying to hide from the comparison.** Both are legitimate purchases. The choice is a values question layered on top of a nearly-identical product decision. --- ## Why Sonnet, not Opus. - Slug: why-sonnet-not-opus - URL: https://ladle.chat/blog/why-sonnet-not-opus - Date: 2026-03-24 - Category: NOTE - Summary: Opus is Anthropic's biggest model. Sonnet is one rung below. For 95% of what people actually ask an AI assistant, Sonnet holds up — and the difference in cost is the difference between a $20 subscription and a $60 one. We get asked why LADLE doesn't run on Opus. Fair question — Opus is Anthropic's largest and most capable model, and if you're paying for AI, "the biggest one" is the intuitive answer. Here's why we don't. Opus costs roughly 5× more per token than Sonnet. At LADLE's average subscriber usage (~800K tokens/month), Opus inference would cost us around $32-38/month per subscriber versus Sonnet's $6-7. Building a $20 product around $32 of variable cost isn't a product; it's a subsidy that runs out. The path most vendors take is to charge more for Opus access. Anthropic itself does — Claude Pro is $20/month for Sonnet with limits; Max is $100-200/month for higher Opus access. That's honest pricing. It's just a different product than what LADLE is trying to be. For LADLE, the constraint is: the meal donation has to be at least $8. Ops has to be paid. That leaves ~$7 for inference. That budget lines up with Sonnet. It doesn't line up with Opus. The second question we get is whether Sonnet is enough. The blunt answer: for what most people ask an AI assistant to do, yes. Drafting emails, summarizing documents, code review on typical PRs, writing help, research assistance, meeting notes, brainstorming — Sonnet handles these at a level that's indistinguishable from Opus for most tasks. Anthropic's own model comparisons back this up; the gap between Sonnet and Opus shows up on the hardest reasoning benchmarks and long complex agentic workflows, not on everyday assistant tasks. Where Opus genuinely pulls ahead: - Multi-step reasoning on hard math, physics, or complex logic problems - Agentic workflows that chain many tool uses across long horizons - Some code tasks on very large codebases For those tasks, if you need them daily, the honest recommendation is Claude Max ($100-200/month) directly through Anthropic. It's the right tool for that job. LADLE isn't the right tool for that job, and pretending it is with a $20 price would be dishonest — either the meals suffer, or the inference budget is unrealistic and the service degrades. There's a subtler thing here. LADLE's design bias is that the model should be quiet and helpful, not the smartest possible entity you have access to. The frontier model race is a fun engineering story. It's not what we're trying to sell. We're selling a competent daily assistant, powered by a well-known model, with a specific and unusual social contract attached. Sonnet is exactly the right tier for that pitch. If Anthropic ships a Sonnet successor that gets materially cheaper, we'll pass the room through — usually as more headroom for heavy users, occasionally as new features. If they ever ship an Opus-tier model at Sonnet-tier prices, we'll switch. Until then, Sonnet is the answer, and it's the honest answer. --- ## The 10-meal minimum: why we chose a floor, not an average. - Slug: the-ten-meal-minimum - URL: https://ladle.chat/blog/the-ten-meal-minimum - Date: 2026-03-14 - Category: NOTE - Summary: Every subscription funds AT LEAST 10 meals. Not 'up to,' not 'on average.' The floor changes the math and the trust in ways an average never could. Two weeks after launch, someone asked why we don't advertise the number bigger. On a good month at ShareTheMeal, $8 buys closer to 11 meals — the price fluctuates a few cents by region and program. Why not say "up to 11"? Because "up to" is the phrase every donation-linked purchase in America uses to hide the fact that it usually pays out less. "Up to 5% of your purchase" almost always turns into 0.4%. "Feeds up to 20 children" almost always turns into 12. The consumer expectation for donation-linked marketing is asterisk, asterisk, asterisk. We refuse to add to that pile. The floor changes what we're promising in a specific way. Every month, if the receipt shows fewer than $8 × subscriber-count paid to WFP, we've broken our word — publicly, in a way anyone can count. That's the promise. Nothing complicated to interpret. The floor also changes what we do when the meal price moves. If WFP's per-meal cost jumps to $0.90 in a specific region, our 10-meal floor still holds because $8 buys 8.9 meals at that price — we round up on our side and pay a hair more per subscriber to keep the count honest. If it drops to $0.70, we don't reduce the meal count; we send $8 anyway and 11.4 meals happen. The floor never floats down. Only up. An average version of this promise would be cheaper to run and harder to trust. Averages get gamed. A bad month can be balanced by a good month. The subscriber who canceled after a bad month never gets their meals back. A floor is unambiguous — every single month, every single subscription, the number is at least ten. The word "minimum" carries weight only if you protect it. So we do a couple of things: The receipt is public with a WFP reference number. You can compare the count we show to the count they show. If they don't match, something is wrong on our end and we owe an explanation. The floor holds even in months where our costs spike. If Anthropic raises API prices and our margin compresses, ops absorbs it. The meal donation is set aside first, at the moment of charge. It doesn't compete with anything else. If we ever can't hit the floor, we say so, refund the affected subscribers' meal portion for that month, and explain what happened. That's the escape hatch. It's designed to feel expensive to use so we never do. Ten meals a month at $0.80 each is not a marketing number. It's a contract. Averages are marketing. Floors are contracts. LADLE runs on a contract. --- ## Why WFP and not other charities. - Slug: why-wfp-not-alternatives - URL: https://ladle.chat/blog/why-wfp-not-alternatives - Date: 2026-03-08 - Category: NOTE - Summary: Any consumer-facing donation model can pick any charity. We picked WFP because the per-meal math is unusually clean, the platform (ShareTheMeal) is auditable, and the scale can absorb whatever LADLE grows into. The first month of LADLE's design, the charity question kept coming up: "why WFP, not GiveDirectly / Malaria Consortium / Doctors Without Borders / your local food bank / a smaller nonprofit that would benefit more from the visibility?" All valid alternatives. Here's the specific reasoning we landed on. **Reason one: the unit economics are unusually legible.** WFP publishes a per-meal cost of $0.80 through their ShareTheMeal partner platform. That number is auditable — it comes from WFP's actual operating costs across the countries they serve, averaged. It fluctuates a few cents by region and program, but it's remarkably stable and it's the number they use publicly. That legibility matters for LADLE. When we tell a subscriber "your $8 buys ten meals," they can independently verify the price. WFP publishes it. ShareTheMeal shows it. If we quoted "your $8 funds mosquito nets" (a real, effective intervention), the per-unit math is less obviously public. Someone would have to trust our conversion. With WFP, we're just repeating WFP's number. **Reason two: scale.** WFP is the largest humanitarian food-assistance organization in the world. In a typical year they reach ~120 million people across 120+ countries. Whatever LADLE grows into — 100,000 subscribers, a million, ten million — WFP can absorb the donation without saturation or misallocation risk. A smaller charity would be a coordination burden at scale we haven't earned yet. WFP's scale is a load-bearing feature of the partnership. **Reason three: receipts.** ShareTheMeal's platform issues machine-readable donation records. That's what makes the LADLE receipt page possible — every month, we publish the WFP reference number for that month's aggregate donation, and anyone can cross-reference it. The alternative would be a monthly statement that says "we donated to X charity" without a linkable proof. That doesn't fit LADLE's honesty-as-marketing pitch. **What about GiveDirectly / effective altruism-recommended charities?** GiveDirectly gives cash to people in extreme poverty; the effectiveness research is strong; we admire the model. Two reasons we didn't pick them: (1) the per-unit ("a household in Kenya received $X") doesn't translate as viscerally as "meal" for consumer marketing without long explanation, and (2) they don't have a partner platform equivalent to ShareTheMeal for aggregated micro-donations with public reference numbers. **What about a smaller charity that would benefit more from the visibility?** We considered this seriously. The counter is: LADLE's donation, at scale, could exceed a small charity's total annual budget. That's a governance and capacity problem — we'd be their biggest donor, which creates strange power dynamics. A charity our size (small) partnering with a nonprofit of enormous scale keeps the relationship uncomplicated. **What about local giving in the country we operate from?** An option, but "meals" globally reads as more universally understood than any specific local intervention. And WFP works everywhere, including in the country we operate from — the same $0.80/meal math applies. **What about a Bill & Melinda Gates Foundation-style outcome fund?** Not accessible at consumer donation scale. Those work at institutional scale with different accountability structures. **Could we ever change partners?** Yes, if a specific reason emerged. If WFP's per-meal math shifted so much it broke the marketing, if ShareTheMeal's platform ceased operating, or if a specific concern arose about WFP's operations that we couldn't answer honestly. We'd announce the switch in the changelog with the reasoning and route future donations to the new partner. The historical receipts wouldn't move. For now: WFP is the answer. It fits the promise, the math, and the scale. --- ## Announcing LADLE. - Slug: launch-announcement - URL: https://ladle.chat/blog/launch-announcement - Date: 2026-03-01 - Category: ANNOUNCEMENT - Summary: We opened subscriptions today. One product, one price, one specific commitment: $8 of every $20 becomes ten meals through the World Food Programme, published monthly with a reference number. LADLE is now live at ladle.chat. Subscriptions are open. The one-line version: it's a general AI assistant powered by Claude Sonnet at $20 a month. Of every $20, $8 is routed to the UN World Food Programme via ShareTheMeal at $0.80 per meal — a minimum of ten meals per subscription per month, with a public receipt. The reason to build this was small and specific. A lot of people are already paying $20 a month for a general AI assistant. The interesting question isn't whether the product is worth $20 (that question is largely settled — frontier models at that price are extremely useful). The interesting question is whether some of that $20 can do a second job on the way out, in a way that's specific and verifiable rather than vague and marketing-flavored. Our answer is receipts. Not "we support hunger relief". Not "a portion of proceeds". A specific transfer amount, a specific meal count at a specific per-meal cost, a specific reference number issued by ShareTheMeal, published on a public URL the day the transfer clears. You can email WFP's donor support with the reference number and verify it yourself. That commitment shapes everything downstream. One plan, one price — we can't invent a Pro tier without fracturing the meal math. No retention offer on cancel — offering a discount to keep you means offering to fund fewer meals. No fundraising deck — the business either works at $20 across enough subscribers or it doesn't, and it's on us to make that math close. There is no free tier. There is no promotional first month. There is a demo at ladle.chat/demo with three free messages, no signup. If it fits your workflow, subscribe. If it doesn't, don't — we'd rather you use ChatGPT or Claude directly than pay us $20 for something you're not using. Every operating decision we make from here reduces to the receipts. If a decision makes the receipts easier to trust, it's easier to make. If a decision makes them harder to trust, it doesn't happen. The whole company is a wrapper around that mechanism. Signup is open now. First receipts publish end of March. ## Canonical resources - Root: https://ladle.chat - Blog index: https://ladle.chat/blog - Blog RSS: https://ladle.chat/blog/rss.xml - Pricing: https://ladle.chat/pricing - Receipts: https://ladle.chat/impact/reports - Changelog: https://ladle.chat/changelog - Roadmap: https://ladle.chat/roadmap - PBC charter: https://ladle.chat/about/pbc-charter - Compact map: https://ladle.chat/llms.txt ## Contact Support: hello@ladle.chat · Press: press@ladle.chat · Privacy: privacy@ladle.chat · Security: security@ladle.chat