메뉴
Use before claiming work is complete, fixed, or passing - runs the deterministic gate (tests, build, lint, types, acceptance) and one light judgment review, with evidence before any claim.
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
Use to build an approved detailed spec task by task in the current session - tests required, single-agent by default, subagents only for large or parallel tasks.
You MUST use this before any creative work - creating features, building components, adding functionality, or modifying behavior. Explores intent, requirements, and design before implementation.
Use when implementation is complete and verified, to integrate the work - merge, PR, or cleanup - after a final evidence check.
Use after a design is approved in brainstorming, to turn it into one layered detailed spec (acceptance criteria, tasks, and how to verify) before touching code.
Use when you hit any bug, test failure, or unexpected behavior - find the root cause before proposing a fix.
Use when starting any conversation - establishes how to find and use decentpowers skills and the workflow they form.
| name | verifying |
| description | Use before claiming work is complete, fixed, or passing - runs the deterministic gate (tests, build, lint, types, acceptance) and one light judgment review, with evidence before any claim. |
Two layers: a deterministic gate that must pass, and one judgment-review pass that catches what machines miss. Deterministic checks are reliable; an LLM grading itself is not - so the machine checks are the gate and the review is advisory.
Announce: "I'm using verifying before calling this done."
No completion claim without fresh verification evidence.
If you have not run the verifying command in this message, you cannot say it passes. Not "should pass", not "looks right" - run it, read the output, then state the result with the evidence.
End-to-end verification.Any failure means the work is not done. Report the actual state with the output, fix, and re-run.
| Claim | Needs | Not enough |
|---|---|---|
| Tests pass | Test output: 0 failures, this run | "should pass", a previous run |
| Build works | Build exit 0 | linter passed |
| Bug fixed | The original symptom now passes a test | code changed, assumed fixed |
| AC met | That AC exercised and observed | "tests pass" in general |
After the gate is green, do one review for what deterministic checks cannot see: logic errors, scope drift (built more or less than the ACs), unclear boundaries, missing error handling. Either review the diff yourself with fresh eyes, or dispatch one reviewer subagent given the diff and the acceptance criteria - not your history.
This is one pass, not a loop. Triage findings: fix Critical and Important now; note Minor for later. If a finding is wrong, push back with evidence - the test or code that proves it. Default the pass for substantial diffs; you may skip it for an [S], single-file, mechanical change.
Gate green and review triaged → invoke finishing to integrate the branch.