메뉴
Root-cause a crash, panic, failing test, or wrong behavior — reproduce, read the real error, trace to the actual cause, then propose the minimal fix. Use when something fails or misbehaves and you need the cause, not a guess. Never edit a test to make a real failure pass.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Drain this repo's .agent/tasks/ queue autonomously and run-to-completion, taking EACH task to a ship-ready bar — claim a 00_todo/ task (`coop tasks claim`), build it, gate it green, self-review it against the .agent/rules KB and every hat, ITERATE until clean, COMMIT it on its own, then `coop tasks done` — without quitting early. Use to "work all the tasks" / drain a backlog / run an unattended sweep.
Cut a versioned release of coop — derive the semver bump from the `## Unreleased` changelog, finalize + annotated-tag it, then (only on explicit confirmation) push the `v*` tag that triggers the public GoReleaser build. Refuses a docs-only no-op release. Use when asked to "cut / ship / tag a release", "publish a release", or "bump the version".
The heavyweight pre-merge review — convene a board of expert hats (correctness, this repo's own rules, security, plus PM/UX/maintainer and any hat the change earns) as parallel reviewers, then synthesize ONE ranked verdict and an ordered plan to fix everything. Reads the project's OWN `.agent/rules/` and runs its OWN gate — no hardcoded laws. Read-only. Works on local changes, a commit, a range, a branch/ref, or a PR. Use before landing anything substantial, or when you want a thorough multi-perspective review.
Execute an approved plan step-by-step, with the repo's gate between steps and no scope creep. Use when implementing a planned change, working a checklist, or the user says "go" / "implement it" / "do the plan". Stops and reports on the first red gate.
Design an opinionated, boring-by-default implementation plan before writing code. Use when a change spans more than one file or context, when the approach isn't obvious, or when asked to plan/design/think through a feature. Produces a small-step, verifiable plan to approve before /work.
Confirm a function, argument, option, type, callback, or CLI flag actually exists with the signature you think — before you call it. Use whenever you're about to use an API you're not certain of, or any time you'd otherwise be guessing. Stops hallucinated calls before they're written.
| name | investigate |
| description | Root-cause a crash, panic, failing test, or wrong behavior — reproduce, read the real error, trace to the actual cause, then propose the minimal fix. Use when something fails or misbehaves and you need the cause, not a guess. Never edit a test to make a real failure pass. |
| argument-hint | <the error, failing test, or symptom> |
| allowed-tools | Read, Grep, Glob, Bash |
Discipline beats guessing. Do not propose a fix until you can point at the line that's wrong and say why. Never edit a test to make a real failure pass.
go test ./pkg -run TestX, the CLI
invocation, the request). No repro → you're guessing./verify-api) — half of "bugs"
are an assumed return shape, arg, or flag that was never real./verify-api it against the real source.$TMPDIR under /var resolving to
/private/var).t.Parallel() without isolation, or an async
effect not made deterministic. Never sleep it away; synchronize on the signal.Cause: <file:line — what's wrong and why> → Fix: <the minimal change> →
Regression test: <what it asserts>. If you can't isolate it, say what you ruled out
and what you'd instrument next — don't ship a guess.