Claude Config Optimizer Skill

Rules and tracking for Claude Code configuration optimization.

1. Config File Location

User's Claude Code config is stored at:

~/ghq/github.com/i9wa4/dotfiles/config/claude/

Key files:

• •settings.json - Main settings
• •CLAUDE.md - Persona and core rules
• •rules/ - Detailed rules (auto-loaded at startup)
• •skills/ - Specialized knowledge (loaded on demand)
• •agents/ - Custom agents
• •commands/ - Slash commands

2. Fetch CHANGELOG

Use gh command to fetch the latest CHANGELOG:

gh api repos/anthropics/claude-code/contents/CHANGELOG.md \
  --jq '.content' | base64 -d > .i9wa4/tmp/claude-code-changelog.md

Then read the file to analyze.

3. CHANGELOG Operations

3.1. Latest Release Summary

1. •Fetch CHANGELOG using the command above
2. •Extract the first ## x.x.x section
3. •Categorize changes into:
   • •New features (Added)
   • •Bug fixes (Fixed)
   • •Improvements (Improved/Changed)
   • •Deprecations (Deprecated)
4. •Present in Japanese with brief explanations

3.2. Version Diff

1. •Ask user for start and end versions
2. •Extract all sections between those versions
3. •Summarize cumulative changes
4. •Highlight breaking changes and deprecations

3.3. Breaking Changes Detection

1. •Search for keywords: Deprecated, Removed, Breaking, Changed
2. •List affected settings and migration paths
3. •Check user's config for affected settings

4. Specification Reference

For detailed questions about Claude Code specifications, features, and usage:

• •YOU MUST: Use the claude-code-guide subagent via Task tool
• •Example queries: hooks, MCP servers, settings, IDE integrations

Task tool with subagent_type: claude-code-guide

5. Settings Categories

6. CLAUDE.md Design Guidelines

• •YOU MUST: Focus only on persona and core guidelines
• •YOU MUST: Split detailed rules into rules/
• •NEVER: Include unnecessary information at startup (reference links, usage details)

7. Configuration Usage

8. Optimization Checklist

Check the following when editing CLAUDE.md:

• • Is the persona definition concise?
• • Are basic rules truly needed at all times?
• • Can detailed explanations be moved to rules/ or skills/?
• • Have reference links been moved to skills/?

9. File Structure Maintenance

When adding/removing files in rules/, skills/, agents/, or commands/:

• •YOU MUST: Update corresponding table in CLAUDE.md section 4
• •YOU MUST: Keep tables alphabetically sorted or logically grouped
• •IMPORTANT: Verify actual files match documentation after changes

10. Optimization Tracking

Last reviewed Claude Code version: v2.1.31 (2026-02-04)

10.1. Applied Optimizations

• • Persona definition minimized
• • Rules split into rules/ directory
• • Skills split into skills/ directory
• • Agents split into agents/ directory
• • Commands split into commands/ directory
• • Reference links moved to skills
• • language setting - set to "follow CLAUDE.md's instructions"
• • mcpToolSearch setting - set to "auto:1"
• • ENABLE_TOOL_SEARCH env - set to "true" (force enable)
• • plansDirectory setting - set to ".i9wa4/plans"

10.2. Pending Considerations

• • SQL schema validation - moved to databricks skill (Section 8)

10.3. Not Adopting

• •showTurnDuration - keep default (show duration)
• •reducedMotionMode - keep default (animations enabled)
• •Keybindings customization - use defaults
• •Completion status reporting - no consumer for this output

10.4. Version Notes

• •v2.1.31: PDF lock fix, sandbox error fix, system prompt improvement for dedicated tools
• •v2.1.30: Read tool pages param for PDFs, /debug command, MCP OAuth pre-configured credentials, reduced motion mode
• •v2.1.29: SessionStart hook saved_hook_context performance fix
• •v2.1.27: --from-pr flag, auto PR-session linking, debug log improvements
• •v2.1.23: spinnerVerbs setting, terminal rendering perf, mTLS/proxy fix
• •v2.1.21: Zenkaku number input (JP IME), auto-compact timing fix, task ID reuse fix
• •v2.1.20: PR review status indicator, --add-dir CLAUDE.md loading, TaskUpdate delete, config backup rotation, background agent permission prompt
• •v2.1.19: CLAUDE_CODE_ENABLE_TASKS env, $ARGUMENTS[0] syntax, skill approval change
• •v2.1.18: Customizable keybindings (/keybindings command)
• •v2.1.16: New task management system with dependency tracking
• •v2.1.15: npm deprecation notice, React Compiler UI perf, MCP stdio timeout fix
• •v2.1.14: Bash history autocomplete, plugin pinning to git SHA, memory fixes
• •v2.1.10: Setup hook, OAuth URL copy shortcut, heredoc template literal fix
• •v2.1.9: plansDirectory, auto:N for MCP, PreToolUse additionalContext
• •v2.1.7: keybindings.json, showTurnDuration, MCP tool search auto mode
• •v2.1.3: Merged slash commands and skills
• •v2.1.0: language setting, skill hot-reload, context: fork

11. Response Format (CHANGELOG)

# Claude Code vX.X.X

## New Features
- [Feature]: [Description]

## Bug Fixes
- [Fix description]

## Improvements
- [Improvement description]

## Deprecated/Removed
- [Affected settings and migration path]

---
Source: https://github.com/anthropics/claude-code

12. site2skill Usage

Convert documentation websites into Claude Agent Skills.

Requirements: Python 3.10+, wget (brew install wget)

uvx --from git+https://github.com/laiso/site2skill site2skill <URL> <SKILL_NAME>

# Example
uvx --from git+https://github.com/laiso/site2skill site2skill https://docs.pay.jp/v1/ payjp

Options:

• •--output, -o - Output directory (default: .claude/skills)
• •--skill-output - Where to save .skill file (default: current directory)
• •--skip-fetch - Skip download (reuse existing files)
• •--clean - Remove temporary files after completion

To update existing skill docs, re-run without --skip-fetch.

13. Reference Links

Official Documentation:

• •Best Practices: https://code.claude.com/docs/en/best-practices.md
• •Docs Map: https://code.claude.com/docs/en/claude_code_docs_map.md

Community Resources:

• •Claude Code config: https://blog.atusy.net/2025/12/15/claude-code-user-config/
• •CLAUDE.md minimization: https://blog.atusy.net/2025/12/17/minimizing-claude-md/
• •site2skill: https://github.com/laiso/site2skill

14. Permission System Reference

14.1. Permission Modes

14.2. Rule Evaluation Order

Rules are evaluated: deny -> ask -> allow. First matching rule wins.

14.3. Bash Wildcard Patterns

{
  "permissions": {
    "allow": [
      "Bash(npm run *)",
      "Bash(git commit *)",
      "Bash(* --version)",
      "Bash(* --help *)"
    ],
    "deny": ["Bash(git push *)"]
  }
}

NOTE: Space before * matters: Bash(ls *) matches ls -la but not lsof.

14.4. Read/Edit Path Patterns

NOTE: * matches single directory, ** matches recursively.

14.5. MCP and Task Permissions

{
  "permissions": {
    "allow": ["mcp__puppeteer__*"],
    "deny": ["Task(Explore)"]
  }
}

14.6. Managed Settings Locations

14.7. Managed-Only Settings

15. Insights-Based Recommendations

Based on usage analysis (55K messages, 4.7K sessions):

15.1. Applied

• •Schema validation for DB operations: See databricks skill Section 8

15.2. Not Applied

• •Completion status reporting: No consumer for this output
• •PreToolUse hook for SQL: Handled by skill guidance instead

References

• •Hooks Reference: https://code.claude.com/docs/en/hooks
• •Permissions Reference: https://code.claude.com/docs/en/permissions
• •Skills Dynamic Context: https://code.claude.com/docs/en/skills#inject-dynamic-context
• •Vercel AGENTS.md Guide: https://vercel.com/blog/agents-md