Audit and optimize your Claude Code configuration. Along the way, claudit caches current Claude Code ecosystem knowledge — settings, plugins, skills, hooks, MCP, performance patterns — so any subsequent agent task can read it via /claudit:knowledge instead of re-fetching docs.
1. The audit (/claudit). A research-backed audit of your Claude Code setup. Claudit first builds expert knowledge from Anthropic's official documentation, then evaluates your configuration against that knowledge — identifying issues, over-engineering, and features you're not using yet.
2. The cache (/claudit:knowledge). A general-purpose Claude Code knowledge primer. Pre-fetched from official docs, version-tracked, TTL-checked. Any agent task that needs current ecosystem knowledge — building a skill, configuring an MCP, authoring CLAUDE.md, debugging hooks — can pull from the cache rather than fetching the same docs again. The audit uses it; you can too.
- Recursive CLAUDE.md discovery: Finds and audits all instruction files — root, subdirectory,
.claude/rules/,CLAUDE.local.md, and@importreferences - Automatic scope detection: Detects whether you're in a project or not, audits everything relevant without prompting
- Research-first architecture: Subagents fetch official Anthropic docs before analysis begins, ensuring the audit knows the full landscape of what's possible — not just what you already have configured
- Over-engineering detection: Identifies where configuration complexity actively hurts performance. Claude does the heavy lifting; verbose CLAUDE.md files, excessive hooks, and MCP sprawl get in the way
- Cross-scope analysis: Detects redundancy between personal and project config, with nuanced cleanup recommendations
- PR-based delivery: Optionally commits changes to a branch and opens a PR with educational inline comments that teach your team Claude Code features
- Persistent memory: Research agents remember findings across audit runs, getting faster and more accurate over time
- Optimized model routing: Research agents run on haiku (fast/cheap for doc fetching), audit agents run on sonnet (consistent quality for config analysis), regardless of your session model
I Audited My Claude Code Config and Found 32% Was Wasted by Damian Galarza
This plugin ships:
- Skills (4):
claudit— the on-demand audit orchestrator. Runs the 6-phase audit (configuration map → research → audit → score → interactive enhancement → optional PR delivery). Accepts an optional[focus-area]argument (MCP,CLAUDE.md,hooks,security, a plugin name, etc.) to deepen the dive on one area without narrowing scope.knowledge— consumer-invoked. Reads the cached Claude Code expert knowledge for the named domain(s) (ecosystem,core-config,optimization, orall); checks freshness and auto-refreshes stale domains. Any agent task can call this to skip fetching docs the cache already has.refresh— consumer-invoked. Force-refreshes the knowledge cache from official Anthropic documentation. Accepts a single domain orall.status— consumer-invoked. Reports the knowledge cache's freshness, TTL, and per-domain coverage. Read-only.
- Commands: none (each skill is invoked via its
/clauditor/claudit:<skill>slash). - Agents (6):
research-core— fetches Claude Code core-configuration documentation (settings, permissions, CLAUDE.md, memory). Dispatched in Phase 1.research-ecosystem— fetches Claude Code ecosystem documentation (MCP, plugins, hooks, skills, subagents). Dispatched in Phase 1.research-optimization— fetches Claude Code performance and over-engineering documentation. Dispatched in Phase 1.audit-global— audits global Claude Code configuration (~/.claude/) against the researched expert knowledge. Dispatched in Phase 2.audit-project— audits project Claude Code configuration (.claude/,CLAUDE.md, subdirectory files, rules) against expert knowledge. Dispatched in Phase 2.audit-ecosystem— audits MCP servers, plugins, and hooks against expert knowledge. Dispatched in Phase 2.
- Hooks: none — claudit installs no Claude Code event hooks, so the hook invariants are vacuously satisfied.
- Opinions: claudit is an on-demand, interactive audit (Phase 4 requires the consumer to select which recommendations to apply). The knowledge cache is plugin-managed at
~/.cache/claudit/with freshness checks at run-time. Decision memory persists to.claude/claudit-decisions.json(project-scoped audit) or~/.cache/claudit/decisions.json(global-only audit), but only when the consumer accepts, rejects, or defers a recommendation. Phase 5 PR delivery requiresgh, runs only when the consumer opts in, and is restricted to project-scoped files (neverCLAUDE.local.mdor~/.claude/).
Boundary conformance (no silent mutation of consumer artefacts): claudit performs no consumer state mutation at plugin-install time. Every write claudit performs — to the knowledge cache, to decision memory, to project files during Phase 4, or to a GitHub PR during Phase 5 — is the result of a slash command the consumer typed and a recommendation the consumer accepted.
/claudit
Or focus on a specific area for a deeper dive:
/claudit MCP
/claudit CLAUDE.md
/claudit hooks
/claudit security
/claudit my-plugin
When a focus area is provided, audit agents still perform their full scope but go deeper on the focus area — more edge cases, line-level detail, and specific fix suggestions. The health report highlights focus-relevant scoring categories with a ◆ marker and presents a consolidated Focus Deep Dive section.
With no arguments, claudit runs a full audit as usual.
The audit runs through 6 phases:
- Configuration Map — Discovers all Claude-related files (instructions, rules, settings, skills, agents, memory, MCP) and presents a structured map
- Build Expert Context — 3 research agents fetch official Anthropic documentation in parallel
- Expert-Informed Audit — Audit agents analyze your global, project, and ecosystem config against expert knowledge (each receives only its relevant slice of the config map)
- Scoring & Synthesis — 6 categories scored with visual health report and ranked recommendations
- Interactive Enhancement — Select which recommendations to apply; changes implemented with before/after scoring
- PR Delivery — Optionally open a PR with educational inline comments, or keep as local edits
Claudit automatically detects scope based on context:
- Inside a git repo → Comprehensive audit (project + global config)
- Outside a git repo → Global-only audit
No prompting needed — just run /claudit and it does the right thing.
- All CLAUDE.md files — root, subdirectory, and
CLAUDE.local.md .claude/rules/*.md— modular rules with YAML frontmatter validation@importreferences — resolution, broken links, circular detection- Cross-file analysis — duplication, conflicts, architecture assessment
- Settings —
.claude/settings.jsonand.claude/settings.local.json - Skills & agents —
.claude/skills/,.claude/agents/ - Memory —
.claude/MEMORY.md
~/.claude/settings.json— settings fields, model config,claudeMdExcludes~/.claude/plugins/installed_plugins.json— plugin versions, install paths, health- User-level instructions —
~/.claude/CLAUDE.md,~/.claude/rules/ - Managed policy —
/Library/Application Support/ClaudeCode/CLAUDE.md(macOS) - Memory —
~/.claude/MEMORY.md
- MCP servers — binary health, duplicate functionality, context cost
- Plugins — structure, legacy patterns, version currency
- Hooks — event types, matchers, timeouts, redundancy
- Detects personal config that duplicates project-specific instructions (recommends removing from personal)
- Preserves general preferences in personal config (informational only)
When fixes are applied, Claudit can open a PR with educational inline comments:
- Each comment explains what changed, why it matters, the Claude Code feature involved, a link to docs, and the score impact
- Only project-scoped files are included (never
CLAUDE.local.mdor~/.claude/files) - Requires
ghCLI — falls back gracefully to local edits if unavailable
| Category | Weight | What It Measures |
|---|---|---|
| Over-Engineering Detection | 20% | Unnecessary complexity, verbosity, redundancy |
| CLAUDE.md Quality | 20% | Structure, conciseness, multi-file quality, imports |
| Security Posture | 15% | Permission hygiene, secrets exposure |
| MCP Configuration | 15% | Server health, tool sprawl |
| Plugin Health | 15% | Version currency, structure patterns |
| Context Efficiency | 15% | Token budget, aggregate instruction size, memory usage |
| Grade | Score | Label |
|---|---|---|
| A+ | 95-100 | Exceptional |
| A | 90-94 | Excellent |
| B | 75-89 | Good |
| C | 60-74 | Fair |
| D | 40-59 | Needs Work |
| F | 0-39 | Critical |
Claudit remembers what you decided about its recommendations. When you accept, reject, or defer a recommendation, that decision is stored in .claude/claudit-decisions.json (project scope) or ~/.cache/claudit/decisions.json (global scope).
On future runs, claudit annotates recommendations with past decisions:
[2] Trim CLAUDE.md redundancy (+15 pts Over-Engineering)
Previously rejected (2026-02-15, acostanzo): "Team onboarding — keeping for junior devs"
⚠ Config changed since decision — recommend re-evaluating
Key principles:
- Context, not constraints — past decisions annotate recommendations but never suppress them
- Staleness detection — decisions are flagged for re-evaluation when config changes, Claude Code updates, score impact shifts, or 90 days pass
- Team-shared — project decisions are committable so teammates see why deviations from best practice were intentional
- Fingerprint matching — recommendations are matched to past decisions via a composite key (category, issue type, file)
Research agents use memory: user to persist findings across runs. The first audit fetches all documentation from scratch. Subsequent runs consult cached knowledge and only update what may have changed — making them faster and more accurate over time.
- Claude Code CLI
- Internet access (for Phase 1 documentation fetching)
ghCLI (optional, for PR delivery)
From the quickstop marketplace:
/plugin install claudit@quickstopOr directly:
claude --plugin-dir /path/to/quickstop/plugins/clauditMIT. See LICENSE.