On this page
Core model
Usage examples
Use concrete repo requests. The skill turns them into bounded work records before editing.
First-run examples
/implementaudit fix the findings in AUDIT.md
/implementaudit < audit.md
/implementaudit --onboard-tools
/implementaudit add tests for the login timeout bug
/implementaudit close the reviewer comments in HANDOFF.md
/goal using /implementaudit, close the findings in AUDIT.md
Natural-language requests are valid when they can be normalized into a bounded work record, for example:
- "implement these findings"
- "act on this audit"
- "close these items"
- "work through this handoff".
--onboard-tools is an explicit optional-tooling path.
It may detect or offer Graphify/ActiveGraph setup, but it still does not make those tools mandatory and does not prove audit correctness.
Governed casual-build intake
/implementaudit add a login page to this app
This is valid only as governed casual-build intake.
Before any mutation, the skill first synthesizes:
- scope
- owner/source
- acceptance criteria
- evidence plan
- rollback or defer path.
Goal synthesis
/implementaudit audit this repo and give me the next best goal
This asks for a bounded plan or handoff.
If already inside a host goal runner, the skill should not emit a nested goal.
Bad examples
/implementaudit do everything- unbounded./implementaudit publish this- publication is a separate gate./implementaudit prove provenance- provenance needs explicit evidence and separate authorization.
Worked trace
| Trace part | Example |
|---|---|
| Finding | README.md describes behavior that no longer matches skills/implementaudit/SKILL.md. |
| Owner/source | skills/implementaudit/SKILL.md defines behavior.README.md is derived public documentation. |
| Smoke A | Read both files, record the mismatch, and run the smallest relevant check such as git diff --check. |
| Countermeasure | Patch the derived documentation, not the canonical behavior source, unless the audit proves the source is wrong. |
| Smoke B | Re-run the check and inspect changed lines for evidence-boundary wording. |
| Closure | Mark the item terminal. Commit only if explicitly authorized. |
Casual-build worked example
- User asks for a repo change such as "make the docs portal generated by CI and prove it is fresh."
- IMPLEMENTAUDIT treats that as governed casual-build intake, not ungated autonomy.
- It synthesizes scope, owner/source, and acceptance criteria.
- It records evidence plan, rollback/defer path, sidecar status, and generated-artifact policy.
- It routes mixed work: new CI surfaces are design work; existing repo behavior is repair work.
- It closes only after checks and final audit evidence support the claim, or it hands off the blocker.