audit-rules

Overview

This skill dispatches the dev-workflow:rules-auditor agent to audit CLAUDE.md rules in a separate context so it starts with a fresh, unbiased perspective.

Process

Step 1: Gather Context

Collect the following before dispatching:

1. Deviation description — from the user's message (e.g., "Claude added animations without asking"). Set to "periodic check" if none provided
2. Global CLAUDE.md path — ~/.claude/CLAUDE.md
3. Project CLAUDE.md path — check for CLAUDE.md in project root (may not exist)
4. Installed plugin behaviors — discover installed plugins and collect their behavioral surface:
   • Glob ~/.claude/plugins/*/*/.claude-plugin/plugin.json to find installed plugins
   • For each plugin found:
     • Read hooks/hooks.json (if exists) — extract hook events, matchers, and script paths
     • Read each hook shell script to extract the behavioral description (first comment line after shebang)
     • Glob skills/*/SKILL.md — extract frontmatter description field from each
   • Compile into a structured summary per plugin:

     Plugin: {name}
     Hooks:
     - {Event}({Matcher}): {description from script comment}
     - ...
     Skills:
     - {name}: "{description from frontmatter}"
     - ...
     

   • If no plugins found: pass "No plugins installed" to agent

Step 2: Dispatch Agent

Use the Agent tool to launch the dev-workflow:rules-auditor agent with all gathered context. Structure the task prompt as:

Audit CLAUDE.md rules with the following inputs:

Deviation description: {description or "periodic check"}
Global CLAUDE.md: ~/.claude/CLAUDE.md
Project CLAUDE.md: {path or "none"}
Installed plugin behaviors:
{compiled plugin behavior summary, or "No plugins installed"}

Step 3: Present Results

When the agent completes:

1. Validate DP count: Parse the Decisions: line for N+M total, count actual DP-\d+ entries in headings. If mismatch, prepend warning to report: ⚠️ 报告声称 {N+M} 个决策点，实际只列了 {actual} 个。以实际条目为准。
2. Present the Rules Review Report returned by the agent
3. Decision Points: Check the agent's return for Decisions: count.
   • If Decisions > 0: read the ## Decisions section from the report
   • For each blocking decision: present to user via AskUserQuestion with options from the decision point
   • For recommended decisions: present as a group via a single AskUserQuestion. Critical: all DP content must be inside the question field — text printed before AskUserQuestion gets visually covered by the question widget. Read each recommended DP's full block (heading + Context + Options + Recommendation) from the report and concatenate them verbatim in the question field, separated by \n---\n. End with: \n\n全部接受推荐，还是逐个审查？
   • If the user does NOT choose to accept all: present each DP individually via separate AskUserQuestion calls. Do not assume any DP is accepted until the user explicitly confirms it
   • Record user choices in conversation (note which option was chosen for each DP)
4. If fix recommendations were made:
   • List each recommendation
   • Ask the user: "Execute these fixes?"
5. If user approves: apply the recommended changes to the CLAUDE.md files
6. If user declines: done

Completion Criteria

• Audit report presented to user
• If fixes recommended: user has approved or declined each recommendation
• If approved: changes applied to CLAUDE.md files