Rule Effectiveness Analysis

Analyze the effectiveness of .claude/rules/ in the current project by cross-referencing rule globs against actual file activity from git history.

Scope (native-first boundary, v4.7.0): this skill owns the one measurement Claude Code has no native equivalent for — does each rule's glob match files the project actually touches. Everything else is native:

• /doctor prompt-audit [path] (v2.1.283+) flags rules/CLAUDE.md written for older models, references to files or commands that no longer exist, and contradictions between files — run it for content quality.
• /skill-doctor (v2.1.252+) and /doctor report per-skill/MCP context cost and unused skills — run them for token budget. Do not re-implement either here.

Step 1: Collect rules inventory

Read all .md files in .claude/rules/. For each:

1. Extract globs: value from YAML frontmatter
2. Count lines of content (excluding frontmatter)
3. Record filename and glob pattern

If a rule has no globs: or paths: frontmatter, classify as always-loaded (loads every session regardless of files touched).

Rules with globs: load eagerly at session start. Rules with paths: + alwaysApply: false load lazily (only when a matching file is touched). Note: paths: must be unquoted CSV — YAML arrays and quoted strings fail silently.

Step 2: Collect file activity from git history

Run: git log --name-only --pretty=format:'' --since='3 months ago' (or configurable period).

Parse output to build:

• session_files: group files by commit date (approximate 1 day = 1 session)
• total_sessions: count distinct dates with commits
• all_files_touched: unique set of all files modified

If fewer than 5 sessions available, warn that results may not be representative and extend to --since='6 months ago'.

Step 3: Cross-reference rules vs activity

For each rule with a glob pattern:

1. Match glob against all_files_touched using bash glob expansion or fnmatch logic
2. Calculate:
   • matched_files: count of unique files that match the glob
   • match_rate: % of sessions where at least 1 file matched the glob
   • token_cost: lines of rule content (proxy for context consumption)

For the project overall:

• covered_files: files that match at least 1 rule glob
• uncovered_files: files touched but matching no rule
• file_coverage: covered / total

Step 4: Classify rules

Step 5: Detect coverage gaps

From uncovered_files, group by directory prefix. Report directories with >5 uncovered files:

SIN COBERTURA:
  src/utils/     — 12 files touched, no rule covers this path
  scripts/       — 8 files touched, no rule covers this path
  migrations/    — 5 files touched, no rule covers this path

For each gap, suggest:

• If directory maps to an existing stack (e.g., migrations/ → supabase), recommend adding stack rule
• If directory is project-specific, recommend creating a custom rule

Step 6: Native content + cost audit (pointer, not re-implemented)

Tell the user to run, in this order, and include the pointers in the report:

1. /doctor prompt-audit .claude/rules — stale-model phrasing, broken paths/commands, contradictions (content quality).
2. /skill-doctor — context cost and usage of skills; /doctor for MCP servers and plugins vs their cost.

For inert rules found in Step 4, the line count is a sufficient proxy — do not compute a token budget table.

Step 7: Generate report

═══ RULE EFFECTIVENESS — {{project}} ═══
Period: last {{N}} months ({{total_sessions}} sessions)
Rules: {{rule_count}} files, {{total_lines}} lines

── ACTIVE (keep) ──
  {{rule.md}}     — {{match_rate}}% match, {{matched_files}} files, {{lines}} lines
  ...

── OCCASIONAL (evaluate) ──
  {{rule.md}}     — {{match_rate}}% match, {{matched_files}} files
  ...

── INERT (candidates for removal) ──
  {{rule.md}}     — {{match_rate}}% match, {{matched_files}} files ← {{reason}}
  ...

── ALWAYS-LOADED ({{total_lines}} lines/session) ──
  {{rule.md}}     — {{lines}} lines, globs: **/*
  ...

── COVERAGE ──
  Files covered by rules:   {{covered}}/{{total}} ({{coverage}}%)
  Gaps: {{gap_dirs}}

── NATIVE FOLLOW-UP ──
  /doctor prompt-audit .claude/rules   → content quality (stale models, broken refs, contradictions)
  /skill-doctor                        → context cost of skills; /doctor for MCP + plugins

── RECOMMENDATIONS ──
1. {{action}} — {{reason}}
2. ...

Step 8: Offer automated fixes

For each recommendation, offer to execute:

• Remove inert rule: rm .claude/rules/{{file}} (with confirmation)
• Narrow glob: edit frontmatter to restrict to actual file patterns
• Add missing rule: create .claude/rules/{{name}}.md with suggested globs
• Split overbroad rule: extract stack-specific content to narrower-glob rule

Only execute with user confirmation. Show diff before applying.