prompt

name	prompt
description	Craft or refactor LLM instructions. Use when using the Agent or Task tool, when the user says "write a prompt", "prompt", "prompt an agent", or any similar request.
metadata	{"user-invocable":true}

Prompt

the user asks for a prompt to be written, refactored, or polished, or pastes an existing prompt or CLAUDE.md asking for improvement.

The output artifact uses a fixed seven-section scaffold. Load assets/template.md and fill it. The procedure below covers detection, inventory, shaping, linting, and emit.

always aim for 80 lines or less in total when it is possible. Keep instructions and sections light and flexible.

Inventory the input

identify what the input contains before writing anything.

in seed mode — gather signal from the conversation. Note the task verb, the domain, any files referenced, constraints stated, register requested, audience implied, and stakes. List ambiguities worth surfacing to the user.

in refactor mode — extract every instruction in the source as a flat list with stable IDs (I-01, I-02, ...). The list is the preservation contract. Truth-conditions are sacred — register changes, the underlying requirements do not. Every ID must end up in at least one section of the output, or it has been dropped.

Shape into the template

load assets/template.md and fill its seven sections from the inventory.

stakes content sits at the close of the Task section. Name what is at risk and why this matters. There is no separate Why section anywhere in the artifact.

a section has no real content — write a one-line gap note that names the absence, for example: "No prior context identified — agent works from conversation signal alone." pad an empty section with filler.

in refactor mode — every extracted instruction ID lands in at least one section. Consolidations are explicit (note which IDs share a slot). An unassigned ID is a dropped instruction.

Lint the draft

load references/eip-anti-patterns.md and read every phrasing in the draft against the patterns it lists.

For each hit, rewrite the phrasing only. The underlying requirement stays. The lint targets register. Truth-conditions are sacred. "You must write tests for every function" becomes "Every function needs a test. Missing tests in a payment system cost more than spurious tests."

writing the natural phrasing first, then checking it against the catalog. The rewrite changes phrasing — never structure, never the underlying instruction. Alignment is what remains after the lint passes; warmth pasted on top of an unaligned prompt does not produce it.

Emit

output only the artifact. No preamble, no postscript, no surrounding commentary, no closing question.

the user supplied a path — write the file directly. they did not — display the artifact inline.

fabricate file paths, scripts, or tools that the surrounding context has not established.

a required assumption was made — note it on the first line of the artifact, not in commentary around it.

in refactor mode and the source contains more than ~15 distinct instructions — show the coverage report (every input ID mapped to a destination section) and ask the user to confirm before writing the file. Smaller refactors and seeds emit without asking.

the user has indicated downstream subagent evaluation — keep section boundaries clear and place the coverage report in a fenced block.

Edge cases

the input is genuinely ambiguous between seed and refactor — default to seed. The assumption note on the first line of the artifact is mandatory.

the user pastes a prompt and asks for one specific tweak — run the full lint anyway, but emit a minimal diff. preserving choices the user did not ask to revisit.

the existing prompt is already aligned and would survive the lint untouched — say so plainly and skip the rewrite. Do not churn for theater.

the source prompt contradicts itself — surface the contradiction with AskUserQuestion. silently resolve it.

the user requests a register the lint would otherwise flag, e.g. terse command style for a one-shot formatter — honor it and name which patterns the requested style relaxes. Domain register wins over universal lint.

References

assets/template.md — the artifact scaffold. Load during shape; the output fills it. references/eip-anti-patterns.md — the anti-pattern catalog with detection hints and rewrite patterns. Load every run. references/eip-principles.md — the principles each anti-pattern derives from. Load when a rewrite needs grounding.