Custom instructions that actually work.

Custom instructions are the paragraph you write at Settings → Behavior that gets appended to every reply's system prompt. Everyone knows they exist. Most people write them once, badly, and never revisit. This is a shame — good custom instructions are the highest-ROI 15 minutes you can spend on your LADLE account.

What custom instructions actually do

They insert a "## Custom instructions from the user" block below the base persona and response style, and above the memory / project / knowledge blocks. The model treats them as high-priority context that should shape every reply.

They do NOT change the model. They don't add capabilities. They don't override safety refusals. They shape voice, format, register, and default behavior on the axes you specify.

The one rule: be specific.

"Be direct" is nearly useless — the base persona already tries to be direct. "Prefer British spelling; cut hedging language; when I paste code, respond with a diff not a rewrite" is three specific instructions that will visibly change every reply.

Write in imperative sentences. Skip the throat-clearing.

Recipe 1: Software engineer

> "I'm a senior backend engineer working mostly in Go and Postgres. Prefer plain code over cleverness. When I paste code, respond with the specific change (diff or block-replacement), not a rewrite. Cite man pages, RFCs, or the actual docs when you assert a behavior. If a decision is a matter of taste, name that and stop — don't relitigate style."

Recipe 2: Designer

> "I lead design at a fintech; Figma-first workflow. Assume I know standard type/spacing/color language. Don't over-explain the fundamentals. When I ask for feedback on a design, structure as: what's working, what's off-brand, three specific improvements ranked by impact. Never soften negative feedback."

Recipe 3: Founder

> "Time is the constraint. Answer in one paragraph unless I ask for depth. When I ask for a decision recommendation, give me one, then the strongest counter-argument, and stop. Don't offer to brainstorm — just recommend. When you're uncertain, say what would resolve the uncertainty."

Recipe 4: Writer

> "Nonfiction essayist. Preserve my voice on edits — never smooth it toward neutral. When I ask for structural feedback, respond in three passes: (1) is the argument load-bearing? (2) where does it wobble? (3) what would you cut? When I ask for line edits, show diffs not rewrites."

Recipe 5: Marketer

> "B2B growth. Numbers over narratives. When I ask for positioning, use the format 'for [audience] who [need], [product] is a [category] that [differentiator]'. When I paste ad copy, give me three alternatives in different registers and mark your favorite. Refuse to use the word 'unlock.'"

Recipe 6: Analyst / finance

> "Investment analyst. Show your working — every number-heavy claim should be reproducible. When I paste data, tell me two non-obvious patterns you see and one plausible confound. Flag every assumption that would flip the answer. Never round without saying so."

Recipe 7: Student

> "3rd-year econ major. When I ask for help understanding a concept, teach me one level deeper than I asked — 'ELI22, then ELI-econ-PhD.' When I paste a problem set, don't solve it for me — walk me through the structure and quiz me on the intuition. Cite the source (textbook chapter, paper, or explanation you'd trust)."

Recipe 8: Educator

> "High-school biology teacher, grades 10-12. When I ask for lesson planning, give me a 45-minute plan with a hook, one activity, an exit ticket, and a 15-word summary I could put on the board. When I ask for feedback on a student's work, structure as: keep this / sharpen this / rethink this — one line each."

Recipe 9: Operations / SRE

> "Ops lead. Runbook-first mindset. When I describe an incident, respond with (1) most-likely cause, (2) one command or check that would confirm, (3) two other candidates to rule out in order. When I ask for process design, always include the failure mode."

Test your instructions

The Kitchen panel on every assistant reply shows the composed system prompt anatomy — including a length count for the custom-instructions block. If your instructions loaded, that number is > 0. If it's zero, something is wrong (usually: they were saved on a different account, or you cleared them and forgot).

You can also test by asking, in a fresh chat, "what are my current custom instructions?" — the assistant should paraphrase them accurately. If it can't, they're not loaded.

What to skip

Skip: instructions about the model's identity ("you are a senior engineer"). The persona work is already done; adding another persona on top usually creates a slightly worse response, not a better one.

Skip: instructions asking the model to be "more truthful" or "less biased." The base training is doing the honest-work; asking for it in your custom instructions signals that you don't trust the base, and the model tends to over-correct.

Skip: long lists of things to never say. The refusal norms are already tight. Custom instructions are for shaping DEFAULT behavior on axes you care about, not for policing edge cases.

Revisit quarterly

The single best habit is a 15-minute quarterly review. Read your current instructions. Are they still accurate to what you actually do? Have your needs shifted? Rewrite. The instructions that served you six months ago are probably not the instructions that would serve you today.