Rules And Skills Audit

Audit policy and task workflows before suggesting consolidation. Shared topics are not proof of duplication: preserve deliberate safety reinforcement, thin entrypoints, distinct audiences, and harness projections that point to a canonical source.

Audit workflow

1. Discover guidance with rg --files --hidden, including nested AGENTS.md, operating files, CLAUDE.md, skill files, harness rules, and instructions under common guidance directories. Respect each file's directory scope. If rg is unavailable, fall back to a scoped find over the guidance roots; note which tool produced the inventory.
2. Build a compact inventory: path, purpose, scope, trigger, dependencies, source-of-truth claims, and notable rules or workflows. Record files read and candidates skipped with reasons.
3. Read the actual files behind links, then compare for:
   • repeated policy or checklists with one clear owner;
   • overlapping triggers or broad workflows that subsume narrow ones;
   • contradictory commands, thresholds, or approval rules;
   • stale, broken, unreachable, or orphaned references;
   • missing boundaries between always-on rules and conditional procedures;
   • Harness projection drift: Check whether harness-specific entrypoints (CLAUDE.md, GEMINI.md, .cursorrules) have accumulated standalone instructions rather than acting as thin adapters pointing to canonical project guidance. Projections must not become divergent second sources of truth.
4. Classify each finding as duplicate, merge candidate, scope/trigger issue, stale/inaccurate, conflict, or improvement. Cite the exact path and heading or line, explain the evidence, and rank impact and risk.
5. Recommend the smallest reversible change. Name the canonical owner, migration map, affected links, and validation required. Do not call word-frequency similarity semantic equivalence, and ignore illustrative fenced examples when checking links.

Always-on vs on-demand separation

Do not classify shared concepts between root rules and skills as duplicates:

• Root guidance (AGENTS.md, OPERATING.md): Must remain thin, universal, and always-on. It owns invariant product boundaries, security guardrails, and routing pointers.
• Skills (skills/<name>/SKILL.md): Loaded on demand when triggered. They own deep operational procedures, domain checklists, and tool execution workflows.

Reject proposals to merge on-demand skill bodies into root instructions or to delete skill procedures on the grounds that root rules "already mention the topic."

Respect directory scoping: Nested guidance files (e.g., frontend/AGENTS.md, scripts/AGENTS.md) apply specifically to their subdirectory tree. Do not consolidate them into the repository root if they contain stack-specific conventions that would pollute context for unrelated tasks.

Report shape

# Guidance audit
## Inventory summary
## Findings
## Keep separate
## Proposed consolidation plan
## No-change conclusion

The report must say which files were reviewed, which were skipped, and whether the evidence supports a change. For each proposed change include its trigger, invariants, exceptions, approval boundary, and post-change checks.

Apply boundary

An audit is read-only and report-first. Do not delete, merge, or rewrite guidance while auditing. A later explicit approval may authorize named findings or a bounded group; apply those edits through skill-authoring, then re-read links, routing, and invariants. Never treat a general request to inspect or optimize as approval for a sweeping rewrite.

Do not broaden this skill into content enrichment, documentation fact-checking, application-code review, provider selection, or worker supervision.