Skip to content
Back to skills

Claude Code Optimizer

ASecurity

Audit and optimize Claude Code installations to current best practices. Use when users ask to check, audit, optimize, or improve their Claude Code setup, configuration, hooks, permissions, or workflow. Triggers on requests mentioning Claude Code settings, CLAUDE.md files, hooks configuration, permissions setup, or performance optimization.

  • 57 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added October 17, 2026
toolsshellbashgitsecurityperformance

Works with

  • claude code
  • mcp

Security analysis

A93/100
  • highPerforms destructive filesystem operations

Pro shows the line behind each finding and how to fix it

Scanned October 17, 2026

npx -y skills add diegosouzapw/awesome-omni-skill --skill claude-code-optimizer --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Claude Code Optimizer?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Claude Code Optimizer
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/diegosouzapw-claude-code-optimizer/badge)](https://www.skillsdirectory.com/skills/diegosouzapw-claude-code-optimizer)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: claude-code-optimizer
description: Audit and optimize Claude Code installations to current best practices. Use when users ask to check, audit, optimize, or improve their Claude Code setup, configuration, hooks, permissions, or workflow. Triggers on requests mentioning Claude Code settings, CLAUDE.md files, hooks configuration, permissions setup, or performance optimization.
---

# 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
```json
{
  "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`

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.