SKILL.md
Evidence before assertion. A claim without a fresh command run behind it is not made.
- Name the claim about to be made: "done", "fixed", "passes", "works".
- Run the commands that would prove it false: the tests, the build, the linter, and the actual behavior that changed.
- When a spec exists at
docs/specs/<slug>.md, gate criterion by criterion: run each criterion's verification command and quote the output that proves it. Then check the diff against the spec's non-goals; a change inside fenced scope fails the claim even with green tests. A spec that still carries an## Open decisionssection is a draft, and a draft proves nothing: no claim can be made against it; route back todrill. - Exercise the changed path end to end, not only its unit tests. If the change has a runtime surface, drive it.
- Read the output. Passing means the output says passing, not that the command exited.
- State what was run and what it showed. If it broke, say it broke, with the output.
Rules:
- Default stance: reject. A claim is false until fresh output proves it; the implementer's word is not evidence.
- Never claim from memory of an earlier run; re-run after the last edit.
- A skipped check is reported as skipped, not implied as passing.
- Review feedback is a claim: verify it before implementing. Unclear feedback is checked, not obeyed.
- Partial verification gets a partial claim: "tests pass; behavior not exercised" is honest, "done" is not.
- A claim no command can prove (good design, not slop, reads well) is a taste claim: route it to the
rubricskill, which judges against written criteria by a fresh subagent. - Fast path (one-sentence plan, no spec file): the gate is that sentence's criterion, not the whole suite. Run only the tests that exercise the change; the changed path still runs end to end.
- Deep mode (the
deepskill is active): exercise the changed path end to end unconditionally and gate every criterion; a partial claim is not accepted as final.