Working with artifacts.

An artifact is anything LADLE lifted out of the conversation into the right-side panel: a long code block, a full HTML document, a 600+-word document, a design. They're automatic — you don't ask for them, LADLE detects them.

Once you understand the trigger, artifacts stop feeling like magic and start feeling like a genuine work surface.

When artifacts trigger

Automatic detection runs on every assistant reply. The heuristic:

- **Code**: any fenced code block of 40 or more lines. Language captured from the fence. - **HTML**: a complete HTML document anywhere in the reply. - **Markdown**: 600+ words of prose IN RESPONSE TO a prompt containing a "write / draft / make / create / compose / generate / outline / plan" verb. This gate prevents casual long replies from being lifted. - **Design**: in a design-mode chat, any HTML the model produced.

The trigger is deterministic. Same input → same output. If your reply is 41 lines of code, you'll get an artifact; if it's 39, you won't. If you asked "explain this" and got a 900-word reply, no artifact; if you asked "draft this" and got the same length, artifact.

The two-pane mental model

Once an artifact exists, the chat continues in the left pane and the artifact lives in the right pane. Regenerate the reply → the panel updates. Ask for a specific edit → the panel updates. The transcript on the left shows a compact "artifact card" where the raw content would have been, so the reading flow stays clean.

This works best when your prompts stay focused on the artifact. Mixing "keep working on this code AND also, unrelated, tell me about the weather" in the same chat produces a messy artifact history. Start a fresh chat for the weather question.

The versioning model — and its current limit

Every artifact has one or more numbered versions. When you regenerate or ask for a change, LADLE APPENDS a new version to the artifact. The panel's version selector lets you browse older versions and compare.

**Current limit**: on the first-turn detection, LADLE creates a new artifact per reply rather than a new version on the same artifact. Meaning: three iterations of a landing page produces three artifact rows, not one artifact with three versions. This is a known gap — versioning-across-turns is on the roadmap. For now, if you want a clean version history, use the Regenerate button on the same reply (that creates versions), not a follow-up message (that creates artifacts).

Keeping the ones that matter

The /artifacts page shows everything you've ever produced. Two habits keep it useful:

**Rename good artifacts.** LADLE picks the title from the first-line pattern of the content (a function name for code, an h1 for markdown, the <title> for HTML/design). Sometimes it's right; often you'll want to rename. The panel header has an inline rename affordance — click the title, edit, hit Enter.

**Star the messages that produced ones you'll want.** If an artifact came from an assistant reply worth remembering, hit the ★ in the actions row. The message shows up in /library grouped by chat. Now you have two indices: /artifacts by output, /library by moment.

When to bail on the artifact model

If you're producing a lot of small snippets (single functions, short paragraphs) that don't hit the 40-line / 600-word threshold, no artifacts land, and your chat becomes long. That's fine — the model still produced valuable output; it just isn't automatically archived. Copy what you want with the ⧉ button on the code block, or drop the whole reply into your own doc.

If you're producing a lot of very long content, artifacts start to feel like a burden — /artifacts has 30 rows all named "Python snippet" and you can't tell them apart. Rename ruthlessly, delete the ones you don't need, and consider projects (which group chats + their artifacts under a shared name).

Sharing artifacts

The artifact panel has Copy and Download. It does NOT have a direct-share primitive — sharing an artifact means sharing the chat that produced it (via the SHARE button in the chat header). The public /share/:slug page will include the artifact's content as a viewable block.

If you want to hand someone JUST the artifact (without the chat context), download it and share the file directly.

The three-artifact rule of thumb

If a chat has produced more than three artifacts, it's probably time to start a new chat. Artifacts are a signal that the assistant made you something substantial; a single chat with a dozen artifacts is usually a signal that you're mixing several projects. Split them.

The moment it clicks

Artifacts stop being noise the day you finish a piece of work, look at the panel, and realize the last thing the assistant made is the exact thing you needed. That happens most often when your prompts got specific enough to trigger a real artifact (long code, real HTML, substantive prose) — and the reason it clicked is because the specification collapsed the model's output space.

Specification is always the lever. Artifacts are just what happens on the other side.