Claude Code Optimizer

Audit and optimize Claude Code installations. Requires Claude Code 2.1.3 or later.

Quick Start

Run the appropriate audit based on user request:

1. Full audit: Invoke audit-orchestrator agent for comprehensive review
2. Specific audit: Invoke specialized agent directly (see Agents below)
3. Quick check: Run inline checks from Audit Checklist

Agents

Invoke via Task tool with appropriate subagent_type:

Agent	Purpose
audit-orchestrator	Coordinates full audit across all areas
config-auditor	Audits settings.json and CLAUDE.md files
hooks-auditor	Audits hook configurations
permissions-auditor	Audits permission patterns
workflow-auditor	Audits commands, agents, and skills
mcp-auditor	Audits MCP server configurations

Agent definitions are in agents/ directory. Install to ~/.claude/agents/ or .claude/agents/.

Audit Checklist

Configuration Files

Check these locations exist and are properly configured:

~/.claude/settings.json          # User settings
.claude/settings.json            # Project settings
CLAUDE.md                        # Project memory (root)
.claude/CLAUDE.md               # Alternative location
CLAUDE.local.md                 # Local overrides (gitignored)
~/.claude/CLAUDE.md             # User-level memory

Settings.json Structure

Valid top-level keys (2.1.3+):

• hooks - Event handlers
• permissions - Tool permissions (allow/deny arrays)
• env - Environment variables
• mcpServers - MCP server configurations
• model - Default model
• theme - UI theme
• respectGitignore - File picker behavior
• language - Response language
• fileSuggestion - Custom @ file search command
• releaseChannel - stable or latest channel toggle

Hook Events (2.1.0+)

Valid event names:

• PreToolUse, PostToolUse, PostToolUseFailure, PermissionRequest
• Notification, UserPromptSubmit
• SessionStart, SessionEnd
• Stop, SubagentStart, SubagentStop, PreCompact

Tool Names for Matchers

• Bash, Read, Write, Edit
• Grep, Glob, Task, TaskOutput
• WebFetch, WebSearch
• NotebookEdit, TodoWrite, KillShell
• LSP (Language Server Protocol, 2.0.74+)
• mcp__* for MCP tools (wildcard: mcp__server__*)

Permission Patterns

{
  "permissions": {
    "allow": ["Read", "Grep", "Glob"],
    "deny": ["Bash(rm -rf:*)"]
  }
}

Wildcard syntax: Bash(git *:*) allows all git commands.

Common Issues

Issue	Fix
Hooks not firing	Check event name spelling (case-sensitive)
Skills not loading	Verify ~/.claude/skills/ or .claude/skills/ path
Permissions ignored	Check allow/deny order; deny takes precedence
CLAUDE.md not read	Must be in project root or .claude/

References

For detailed audit procedures, see:

• references/config-guide.md - Configuration deep-dive
• references/hooks-guide.md - Hook patterns and examples
• references/permissions-guide.md - Permission strategies
• references/security-guide.md - Security audit procedures
• references/troubleshooting.md - Common problems and solutions

Version Compatibility

This skill targets Claude Code 2.1.3+ features:

• Skill hot-reload
• Forked sub-agent context (context: fork)
• Hook frontmatter in skills/agents with once: true option
• YAML-style allowed-tools
• Unreachable permission rule detection
• MCP wildcard permissions (mcp__server__*)
• Named sessions (/rename, /resume <name>)
• Release channel configuration

Check version: claude --version Update: claude update