Skills tool

A skill earns its place only by stopping a mistake the model makes without it, in the fewest words that takes. Produce a skill whose every line changes what the model does, with each rule carrying its boundary reason. Avoid restating default behavior, hiding the one rule that changes anything, and an order stripped of its boundary reason.

When to use

• Use it when creating, changing or sizing a skill, agent, hook or skill script.
• When a stage borrows it, hand control back once the edit is done. Otherwise it owns the turn.
• Skip it for a one-time fix, one project's habit, or a limit a regex can check. A commit, CLAUDE.md or the verifier holds those, and a skill is for a call that needs judgment.

The loop

1. Catch the mistake. Write one prompt per references/pressure-scenarios.md whose answer shows the mistake as one observable symptom.
2. Watch it happen. Read benchmarks/results/ and any earlier pressure.mjs answers directory for the case first, because a stored result that covers the edited file and was made after its last edit stands in for the run. When the user asks for a run, or none covers the file, run scripts/pressure.mjs --prompt <file> --cells <cells> --plugin-dir <clone> on the cells --cells-for <edited file> prints, else the session's own, and read only the without answers now.
3. Choose the home. Place the fix per references/where-a-fix-lives.md. Write prose in the shape below, with style per references/instruction-style.md and register per references/wording.md.
4. Watch it stop. Rerun pressure.mjs on the same cells only under step 2's condition (the user asks, or no stored result covers the file), and read the with answers. Rerun with --full on the verdict INCONCLUSIVE from --grade, and with --contrast-env when one setting alone is in question; when the without run also passed, the case shows nothing, so harden it until that run fails.
5. Close each new excuse. Apply references/plugging-holes.md to each justification the with run still produced, then rerun all cases. You are done when a full rerun adds nothing to its tables.
6. Verify. Run the check the caller's rules assign to this edit, and only the proof it names when the caller reserves the full check to its lead. On a red check, fix the structure and add no exemption, lock or allowlist entry, which only hides the failure.
7. Judge blind. When step 4's read was a close call, judge the stored pre-edit and edited with answers blind per references/blind-eval.md, and run its clones only when the user asks.
8. Name the evidence. Commit a skill or agent wording change with an Evidence: benchmarks/results/<file> or Evidence: none (<reason>) trailer, because npm run check fails the branch without one.

The shape

Form

• Move bulk material (a template, a worked example, a checklist over 20 lines) to references/, and let the body name when to open it.
• Aim the body at 2,000 tokens (bytes after the frontmatter / 4) and the description at 200 characters; the verifier fails 2,500 tokens and 250 characters.
• A reference holds one topic and links to no other reference; over 100 lines it opens with a contents list linking each section.
• Put text handed to a delegate beside SKILL.md as <role>-prompt.md, with a References row naming the dispatching step. A delegate is a separate agent the skill starts; references/ holds only what the skill reads itself.

Red flags

References

Judgment

• A rule the without run broke outranks a rule that only reads well.
• Revert an edit that has neither a run nor a stored result made after its last edit and covering its file, and do not keep it as a draft.
• When a skill runs long, drop its weakest rule and keep every rule's reason clause.