| name | review-rules |
| description | Review the core instruction layer — CLAUDE.md, .claude/rules/*.md, and RIPER/CAT mode definitions — for size, staleness, cross-reference integrity, alignment with active hooks, and redundancy. Creates board tasks for every finding. |
| disable-model-invocation | true |
| user-invocable | true |
Rules & Core Instructions Review Protocol
Reviews the always-loaded instruction layer that governs every Claude Code session in this repo. Scope:
CLAUDE.md (project root)
.claude/rules/*.md
.claude/settings.json hooks config (only to verify rule-hook alignment)
Part of the review-* skill family. Cites ../_shared/review-checklist.md and ../_shared/anthropic-configuration-guide.md.
Owner: docs-manager agent.
When to invoke
- CLAUDE.md approaching 150 lines (lean limit) or > 200 lines
- After significant process changes (new hooks, new RIPER modes, new deployment patterns)
- Quarterly audit
- When new team members or agents repeatedly make mistakes the rules should prevent
Output contract
Board tasks only. Lean summary to main agent.
Step 1 — Inventory
- Read
CLAUDE.md. Record line count, section list, referenced rules/skills/docs/agents.
- Read every
.claude/rules/*.md. Record line count, topic, referenced files.
- Read
.claude/settings.json hooks block. Record which hooks exist and what they enforce.
Step 2 — Size discipline
| Artifact | Target | Hard limit |
|---|
| CLAUDE.md | < 150 lines | 200 lines |
| Individual rule | < 200 lines | 300 lines |
Flag bloat. Recommend:
- CLAUDE.md bloat → push detail into a new rule or existing rule
- Rule bloat → split by sub-topic or move deep reference to a skill / doc
Step 3 — Staleness
For each rule and for CLAUDE.md, check:
- Deprecated infrastructure references (Supabase Edge Functions, old table names, retired services)
- Domain names and URLs still current ( variants, MIG host IDs, LB IP
<YOUR_LB_IP>)
- Model IDs match current Anthropic lineup (Opus 4.8, Sonnet 4.6, Haiku 4.5)
- Command references valid (
Enter C -push-dev, Enter C -push, Enter C -all)
- Stripe API version current (
2025-08-27.basil per CLAUDE.md)
- Referenced docs exist (every
[link](path) resolves)
- Referenced agents exist in
.claude/agents/
- Referenced skills exist in
.claude/skills/
Step 4 — Cross-reference integrity
- Every
[text](relative/path.md) in CLAUDE.md and rules → file exists (Glob)
- Every rule file → referenced from CLAUDE.md at least once (orphans are dead weight)
- Every rule topic → has an invocation pattern documented
Step 5 — Rule-hook alignment
Match the written rules against the behavior enforced by .claude/settings.json:
- Does
agent-board.md describe hooks that actually exist?
- Does
riper-cat.md describe auto-transitions that the Stop hook actually enforces?
- Does CLAUDE.md reference hooks that are live?
- Are there hooks in settings.json that no rule explains to humans?
Flag misalignments.
Step 6 — RIPER-CAT accuracy
Specifically for riper-cat.md:
- All 8 modes (R/I/P/E/RE/C/AI/T) described and mapped to the Enter signal
- Auto-transitions described match
/plan and /review skill behavior
- COMMIT flags (
-commit, -push, -push-dev, -all, -return) match rules/deployment.md
- Agent assignment rules cross-check with
/plan skill + .claude/agents/ roster
Step 7 — Redundancy
For the ruleset as a whole:
- Content duplicated across rules (flag for consolidation)
- Content duplicated between CLAUDE.md and a rule (rule wins — pull from CLAUDE.md)
- Two rules that should be merged (near-duplicate scope)
- One rule that should be split (two unrelated topics joined)
Step 8 — Missing rules
Scan recent memory (MEMORY.md), recent commits, and recent incident logs for recurring patterns not yet codified as rules. Candidates:
- Ops patterns that caused repeat incidents
- Coding rules recently learned the hard way
- Deployment patterns not yet documented
Flag as "Codify pattern [X] as a rule".
Step 9 — File findings to the board
Create parent tasks per severity + subtasks on project <YOUR_AUDIT_PROJECT_ID>. Assigned agent: docs-manager.
Step 10 — Summary
Return ≤ 100 words:
RULES REVIEW COMPLETE
- CLAUDE.md: [N] lines (target <150)
- [X] rule files reviewed
- [C/H/M/L] findings per severity → board parent tasks
- [MR] missing-rule candidates
- [RA] redundancy/alignment issues
Anti-patterns to flag
- CLAUDE.md growing past 200 lines (must reduce)
- Rules that contradict each other
- Rule behavior described in CLAUDE.md but not enforced by hooks (drift risk)
- Dead links in CLAUDE.md or rules
- References to retired model IDs or deprecated commands
- Rules that have never been referenced (orphans)
Tools used
- Read, Glob, Grep
- Optionally: read
.claude/settings.json to verify hook alignment
- Board MCP (create_task, update_task, log_activity)