| name | rulesgen-harness-update |
| description | Use when a coding or testing session reveals that a Rulesgen harness file (rule, skill, agent, or doc) is inaccurate, incomplete, or blocking correct work. Requires human approval before any harness file is changed. |
Rulesgen Harness Update Flow
Harness files (.claude/rules/, .claude/skills/, .claude/agents/, .claude/CLAUDE.local.md, root CLAUDE.md) are the operating instructions for all agents working on this project. They must not be changed autonomously — every update requires explicit human approval.
Trigger conditions
Propose an update when:
- A rule contradicts the current code (e.g. lists a Flask convention while the code is FastAPI).
- A skill points at a command that no longer works (e.g. references
python -m pytest while the project uses uv run pytest).
- A path/glob in a rule frontmatter does not match the actual directory layout.
- An agent's tool list is too broad or too narrow for its actual responsibility.
- A
CLAUDE.md "project facts" item is stale (version bump, dependency change, new module).
- A new common failure mode has appeared that no existing skill covers.
Workflow
- Identify the specific harness file(s) and the exact lines that are wrong, missing, or misleading.
- Quote the current text and propose the replacement text side-by-side.
- Explain the trigger: what session, what failure, what consequence.
- Stop and request human approval — do not edit the harness file in the same turn.
- After approval, apply the change in a single, minimal commit (
docs: Conventional Commit type).
- Cross-check: a change to one file may need a mirrored change in another (e.g. updating
rulesgen-test-selection may require an update to testing-contract.md).
Do not
- Do not auto-update the harness during a feature implementation session.
- Do not silently add new agents, skills, or rules without approval.
- Do not edit
.claude/settings.json permissions to "make a command work" — escalate the underlying need instead.
