claudemd-maintainer
Context-aware guidance for maintaining and improving CLAUDE.md files. Use when editing CLAUDE.md, discussing documentation structure for AI assistants, or optimizing project instructions.
Best use case
claudemd-maintainer is best used when you need a repeatable AI agent workflow instead of a one-off prompt.
Context-aware guidance for maintaining and improving CLAUDE.md files. Use when editing CLAUDE.md, discussing documentation structure for AI assistants, or optimizing project instructions.
Teams using claudemd-maintainer should expect a more consistent output, faster repeated execution, less prompt rewriting.
When to use this skill
- You want a reusable workflow that can be run more than once with consistent structure.
When not to use this skill
- You only need a quick one-off answer and do not need a reusable workflow.
- You cannot install or maintain the underlying files, dependencies, or repository context.
Installation
Claude Code / Cursor / Codex
Manual Installation
- Download SKILL.md from GitHub
- Place it in
.claude/skills/claudemd-maintainer/SKILL.mdinside your project - Restart your AI agent — it will auto-discover the skill
How claudemd-maintainer Compares
| Feature / Agent | claudemd-maintainer | Standard Approach |
|---|---|---|
| Platform Support | Not specified | Limited / Varies |
| Context Awareness | High | Baseline |
| Installation Complexity | Unknown | N/A |
Frequently Asked Questions
What does this skill do?
Context-aware guidance for maintaining and improving CLAUDE.md files. Use when editing CLAUDE.md, discussing documentation structure for AI assistants, or optimizing project instructions.
Where can I find the source code?
You can find the source code on GitHub using the link provided at the top of the page.
SKILL.md Source
# CLAUDE.md Maintainer (Smart Router)
## Purpose
Context-aware guidance for maintaining and improving CLAUDE.md files. Helps ensure the file stays effective, concise, and follows best practices for LLM instruction files.
## When Auto-Activated
- Editing or discussing CLAUDE.md
- Keywords: claude.md, project instructions, onboarding claude, context file
- Discussing documentation structure for AI assistants
## 🚨 CRITICAL RULES
1. **Keep under 300 lines** - Research shows LLMs handle ~150-200 instructions reliably; performance degrades with more
2. **Never auto-generate** - Manually craft each line; auto-generated content often contains noise
3. **Universal applicability only** - Remove task-specific or narrow-scope guidance
4. **File references over code snippets** - Code embeds become outdated; paths stay accurate
## 📊 Effective CLAUDE.md Principles
### What To Include (WHAT, WHY, HOW)
| Category | Include | Avoid |
|----------|---------|-------|
| **Stack** | Technologies, architecture overview | Exhaustive dependency lists |
| **Critical Rules** | Must-follow behaviors (consolidated) | Duplicate rules across sections |
| **Quick Commands** | Essential build/run commands | Full command reference |
| **File References** | Paths to detailed guides | Embedded code that can outdated |
| **Common Mistakes** | Documented actual failures | Hypothetical warnings |
### What To Exclude
- **Code style guidelines** → Use linters instead (SwiftFormat, ESLint, etc.)
- **Exhaustive command references** → Point to tool documentation
- **Task-specific instructions** → Put in skills or separate docs
- **Code snippets** → Use `file:line` references instead
## 🎯 Progressive Disclosure Pattern
```
Level 1: CLAUDE.md (~100-150 lines ideal, max 300)
├─ Critical rules (must-follow behaviors)
├─ Quick start (essential commands only)
├─ High-level architecture
└─ Pointers to Level 2 and 3
Level 2: Skills (.claude/skills/)
├─ Domain-specific quick references
├─ Decision trees and workflows
└─ "→ Routes to" comprehensive docs
Level 3: Specialized Guides
├─ Complete technical documentation
├─ Full examples and edge cases
└─ Troubleshooting guides
```
## 📋 Quick Audit Checklist
When reviewing CLAUDE.md:
- [ ] **Line count under 300?** (`wc -l CLAUDE.md`)
- [ ] **No duplicate rules?** (Search for repeated concepts)
- [ ] **Code snippets minimized?** (Replace with file path references)
- [ ] **Narrow-scope items removed?** (Move to skills)
- [ ] **Critical rules consolidated?** (Single authoritative location)
- [ ] **File references current?** (Paths still valid)
- [ ] **Common mistakes documented?** (Actual failures, not hypotheticals)
## 🔄 Optimization Workflow
### Reducing Line Count
1. **Find duplicates**: Search for repeated concepts
```bash
# Find potential duplicates
grep -n "NEVER" CLAUDE.md | head -20
```
2. **Consolidate rules**: Move all critical rules to one section
3. **Replace code with references**:
```markdown
# ❌ Before (takes 10+ lines)
### Typography
```swift
AnytypeText("Title", style: .uxTitle1Semibold)
AnytypeText("Body", style: .bodyRegular)
```
# ✅ After (1 line)
**Typography** → `path/to/TYPOGRAPHY_MAPPING.md`
```
4. **Move narrow guidance to skills**: If it applies to < 20% of tasks, it's a skill
### Adding New Guidance
**Before adding, ask:**
1. Does this apply to most tasks? (If no → skill or specialized doc)
2. Is there existing guidance? (If yes → consolidate, don't duplicate)
3. Can a linter/tool enforce this? (If yes → use the tool instead)
4. Will this code example outdated quickly? (If yes → use file reference)
## ⚠️ Common Mistakes
### Accumulating Hotfixes
```markdown
# ❌ WRONG - Adding one-off rules
### Special Note (2025-01-15)
Remember to always check X when doing Y...
# ✅ CORRECT - Add to appropriate skill or remove
```
### Duplicating Rules
```markdown
# ❌ WRONG - Same rule in 3 places
## Critical Rules: NEVER add AI signatures
## Pre-Commit: NO AI signatures
## PR Format: No "Generated with Claude"
# ✅ CORRECT - Single consolidated rule
## Critical Rules
2. **NEVER add AI signatures anywhere** - No Co-Authored-By, no emoji signatures
```
### Embedding Code That Outdates
```markdown
# ❌ WRONG - Code will outdated
```swift
Image(asset: .X32.qrCode) // What if asset name changes?
```
# ✅ CORRECT - File reference stays accurate
**Icons** → `Modules/Assets/.../ImageAsset.swift:45`
```
## 📚 Research & Best Practices
**Source**: Based on industry practices and LLM behavior research
### Key Findings
- LLMs handle ~150-200 instructions reliably
- Performance degrades with additional instructions
- Irrelevant context may be ignored entirely (Claude adds "this context may or may not be relevant")
- Code snippets in docs become maintenance burden
- Progressive disclosure reduces context overhead
### Recommended Metrics
| Metric | Target | Max |
|--------|--------|-----|
| Total lines | 100-150 | 300 |
| Code blocks | 2-4 | 6 |
| Critical rules | 3-5 | 10 |
| Sections | 6-8 | 12 |
## 🔗 Related
- `.claude/skills/README.md` - Skills system overview
- `.claude/skills/skills-manager/SKILL.md` - Managing the skills system
- Progressive disclosure architecture documentation
---
**Navigation**: This skill helps maintain CLAUDE.md quality. For skills system management, see `skills-manager`. For adding new skills, see `.claude/skills/README.md`.Related Skills
afk-maintainer
Governance skill for the AFK framework. Enforces coding principles, DX-first design, extensibility patterns, production safety, and library quality standards across every contribution. Use this skill when reviewing PRs, triaging issues,planning releases, or making architectural decisions for AFK.
bgo
Automates the complete Blender build-go workflow, from building and packaging your extension/add-on to removing old versions, installing, enabling, and launching Blender for quick testing and iteration.
Buffer Overflow Payload Generator
Generates a buffer overflow attack payload with a specific stack layout (padding, return address, NOP sled, shellcode) and saves it to a file.
browser-testing
Use when testing web applications, debugging browser console errors, automating form interactions, or verifying UI implementations. Load for localhost testing, authenticated app testing (Gmail, Notion), or recording demo GIFs. Requires Chrome extension 1.0.36+, Claude Code 2.0.73+, paid plan.
browser-fetch
Delegate browser automation to a lightweight subagent (Haiku) to reduce context consumption. Also provides web clipping (HTML→Markdown) via clipper.
Browser Automation Expert
浏览器自动化与网页测试专家。支持基于 MCP 工具(Puppeteer/Playwright)的实时交互,以及基于 Python 脚本的复杂自动化流实现。
bronze-layer-setup
End-to-end Bronze layer creation for testing and demos. Creates table DDLs, generates fake data with Faker, copies from existing sources, and configures Asset Bundle jobs. Covers Unity Catalog compliance, Change Data Feed, automatic liquid clustering, and governance metadata. Use when setting up Bronze layer tables, creating test/demo data, rapid prototyping Medallion Architecture, or bootstrapping a new Databricks project. For Faker-specific patterns (corruption rates, function signatures, provider examples), load the faker-data-generation skill.
brand-identity
Provides the single source of truth for brand guidelines, design tokens, technology choices, and voice/tone. Use this skill whenever generating UI components, styling applications, writing copy, or creating user-facing assets to ensure brand consistency.
brainstorming
Use when creating or developing anything, before writing code or implementation plans - refines rough ideas into fully-formed designs through structured Socratic questioning, alternative exploration, and incremental validation
boxlog-frontend-design
BoxLog専用のフロントエンドデザインスキル。「装飾のない基本体験」を実現するためのUI設計ガイドライン。STYLE_GUIDE.mdを補完し、フォント・アニメーション・デザイン判断基準を提供。
bounty-hunter
Find, evaluate, and submit online bounties and hackathons for prize money. Use when user mentions "bounties", "hackathon", "earn money", "Superteam Earn", "prize money", "submissions", "freelance bounties", or asks to find paid opportunities. Covers discovery, eligibility filtering, content drafting, and submission workflows.
bootstrap-phase-workflow
Integrate the vibe/mature phase workflow into a project