Bridle Harness · CLAUDE.md example

Drop a CLAUDE.md (or equivalent .cursorrules / copilot-instructions.md) at your repo root. This is the System Context component of the Harness Five-Pack — see three-leaps.md §9.1.

Project context

This project follows the Bridle / Three Leaps AI coding governance methodology. Before generating any code, you (the AI agent) must:

1. Read the relevant module.yaml for the module being modified.
2. Check the intent.yaml for the desired-state declaration.
3. Verify the action’s reversibility tier (R0–R5):
   • R0–R1: you may execute directly
   • R2: you may execute, must write to docs/agent-decisions/
   • R3: you propose a PR, do not auto-merge — human review required
   • R4: stop and ask the human; do not proceed
   • R5: refuse — this is a permanent red line (financial / physical / irreversible user impact)

Boundaries

• Do not edit modules outside the one named in the user’s prompt unless explicitly asked.
• Do not modify module.yaml lifecycle.state fields without a corresponding intent change and a state-transition rationale.
• Do not humans-in-the-loop bypass — never use --no-verify or skip CI.
• Do not delete files in archive/ or docs/agent-decisions/ (append-only audit trail).

Quality gates

Every PR you open must satisfy:

• lint + type + unit-test + contract-test + sec-scan all green
• New code coverage ≥ 80%
• No new boundary violations (run archtest)
• intent.yaml field updates have matching verifier updates

Decision audit

Every state-changing action you take is logged to docs/agent-decisions/<module-name>.md with:

• timestamp
• triggering signal
• action taken
• reversibility tier
• rollback procedure

This is mandatory under the AI Decision Auditor role (three-leaps.md §13.3).

Anti-patterns to avoid

Anti-pattern	Why it’s bad
Filling in signals.* values manually	Must be mechanically collected (three-leaps.md §5.4)
Promoting experimental → asset without candidate step	Skips the candidate verification gate
Editing module.yaml outside the module’s own directory	Cross-domain manifest writes are R3
Adding fields to intent.yaml without binding a verifier	Anti-pattern A2: “intent ≠ execution”

When in doubt

Read these in order:

1. three-leaps.md §10 (autonomy gradient) — to know what tier this action is
2. three-leaps.md §12 (anti-patterns) — to check you’re not stepping on a known trap
3. The module’s own module.yaml audit.reversibility_max — caps what you may auto-act on

If still uncertain, stop and ask the human. Asking is always cheaper than rolling back.