claude-skills-supercharged

KO | EN

gitlite — search

#javascript #python #go #golang #machine-learning #rust #deep-learning #linux #docker #macos #awesome #kubernetes #security #openai

claude-skills-supercharged

(★ 41)

A "supercharged" implementation of Claude Code Skills – using Haiku prompt analysis/critical skill scoring and skill auto-injection for friction-free, context-driven workflows.

claude-skills-supercharged 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Claude Skills Supercharged

[![Tests](https://img.shields.io/badge/tests-passing-brightgreen)](https://github.com/jefflester/claude-skills-supercharged)
[![TypeScript](https://img.shields.io/badge/TypeScript-5.3-blue)](https://www.typescriptlang.org/)
[![License](https://img.shields.io/badge/license-MIT-blue.svg)](https://raw.githubusercontent.com/jefflester/claude-skills-supercharged/main/LICENSE)
[![Claude
Code](https://img.shields.io/badge/Claude%20Code-Skills-purple)](https://code.claude.com/docs/en/skills)

**AI-powered skill auto-injection system for [Claude Code
Skills](https://code.claude.com/docs/en/skills)**

**Stop manually invoking skills.** Let AI detect what you need and inject it
automatically with 95%+ accuracy. The system analyzes your prompts using Claude
Haiku 4.5, assigns confidence scores, and intelligently loads the right
context—every time.

> 💡 **Credit:** Base reference architecture from
> [@diet103](https://github.com/diet103)'s
> [claude-code-infrastructure-showcase](https://github.com/diet103/claude-code-infrastructure-showcase)

---

## 📊 Performance at a Glance

| Metric | Value |
|--------|-------|
| ✅ Test Coverage | **All tests passing** |
| ⚡ Response Time | **<10ms** (cached) / ~200ms (first call) |
| 💰 Monthly Cost | **$1-2** @ 100 prompts/day |
| 🎯 Accuracy | **95%+** skill detection rate |

---

## 🎯 Why This?

**The Problem:** Manually loading skills is tedious, error-prone, and breaks
your flow. You forget which skills exist, waste time invoking them, and lose
context between conversations.

**The Solution:** This system uses AI to automatically:
- 🔍 **Analyze** your prompt intent with Claude Haiku 4.5
- 📊 **Score** each skill's relevance (0.0-1.0 confidence)
- 🚀 **Inject** high-confidence skills (>0.65) into context automatically
- 💾 **Cache** results for 1 hour to reduce costs
- 🔗 **Track** loaded skills per session (no duplicates)

**Before:** "Let me manually load python-best-practices... wait, was it
python-style? Let me check..."

**After:** Just say "help me write Python code" and the right skills load
automatically. 🎉

---

## 📑 Table of Contents

- [Features](#-features)
- [Quick Start](#-quick-start)
- [How It Works](#-how-it-works)
- [Included Skills](#-included-skills)
- [Architecture](#-architecture)
- [Creating Custom Skills](#%EF%B8%8F-creating-custom-skills)
- [Keeping Skills Updated](#-keeping-skills-updated)
- [Configuration](#%EF%B8%8F-configuration)
- [Testing](#-testing)
- [Troubleshooting](#-troubleshooting)
- [Documentation](#-documentation)
- [FAQ](#-faq)

---

## ✨ Features

- 🤖 **AI-Powered Intent Analysis** — Claude Haiku 4.5 analyzes your prompts to
  detect relevant skills with 0.0-1.0 confidence scoring
- ⚡ **Automatic Injection** — High-confidence skills (>0.65) automatically
  loaded into context
- 💾 **Smart Caching** — 1-hour TTL reduces API costs (~$1-2/month) and improves
  response time (<10ms cached)
- 🔄 **Session Tracking** — Skills only injected once per conversation, no
  duplicates
- 🔗 **Bidirectional Affinity** — Related skills automatically loaded together
- 📚 **Progressive Disclosure** — Main skill files under 500 lines, detailed
  content in resources
- 🛡️ **Guardrail Skills** — Enforce critical best practices (e.g., API
  security, input validation)
- ✅ **Comprehensive Testing** — Full coverage of the injection
  pipeline

## 🚀 Quick Start

### 1. Clone the Repository

```bash
git clone https://github.com/jefflester/claude-skills-supercharged.git
cd claude-skills-supercharged
```

### 2. Install Dependencies

```bash
cd .claude/hooks
npm install
```

### 3. Configure API Key

```bash
cp .env.example .env
# Edit .env and add: ANTHROPIC_API_KEY=sk-ant-your-key-here
```

Get your API key from: https://console.anthropic.com/

### 4. Test the System

```bash
# Test skill activation
echo '{"session_id":"test","prompt":"I need to write Python code"}' | \
  npx tsx skill-activation-prompt.ts
```

### 5. (Optional) Set Up Pre-commit Hooks

```bash
# Install pre-commit
pip install pre-commit  # or: brew install pre-commit

# Install git hooks
pre-commit install

# Run on all files (optional)
pre-commit run --all-files
```

### 6. Use with Claude Code

The system automatically activates when you use Claude Code in this project. Try
these prompts:

- "Help me write a Python function" → Loads `python-best-practices`
- "I need to add a REST API endpoint" → Loads `api-security`
- "How do I write a good commit message?" → Loads `git-workflow`

You'll see a banner showing which skills were loaded:

```
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📚 AUTO-LOADED SKILLS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

<skill name="python-best-practices">
[Skill content automatically injected...]
</skill>

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
```

## 🔧 How It Works

The skills system uses a **7-stage injection pipeline**:

```
┌─────────────────────────────────────────────────────────────┐
│ 1.  User Prompt                                             │
│    "Help me write Python code"                              │
└────────────────────┬────────────────────────────────────────┘
                     ↓
┌─────────────────────────────────────────────────────────────┐
│ 2.  Intent Analysis                                         │
│    • AI Analysis (Claude Haiku 4.5)                         │
│    • Keyword Fallback (if API unavailable)                  │
│    • Cache Check (MD5 hash, 1-hour TTL)                     │
└────────────────────┬────────────────────────────────────────┘
                     ↓
┌─────────────────────────────────────────────────────────────┐
│ 3. Confidence Scoring                                       │
│    python-best-practices: 0.92     Required (>0.65)         │
│    git-workflow: 0.58              Suggested (0.50-0.65)    │
│    api-security: 0.12              Ignored (<0.50)          │
└────────────────────┬────────────────────────────────────────┘
                     ↓
┌─────────────────────────────────────────────────────────────┐
│ 4.  Skill Filtration                                        │
│    • Remove already-loaded skills                           │
│    • Apply 2-skill injection limit                          │
│    • Promote suggested skills if slots available            │
└────────────────────┬────────────────────────────────────────┘
                     ↓
┌─────────────────────────────────────────────────────────────┐
│ 5.  Affinity Injection                                      │
│    • Load bidirectionally-related skills                    │
│    • Free of slot cost (bonus injections)                   │
└────────────────────┬────────────────────────────────────────┘
                     ↓
┌─────────────────────────────────────────────────────────────┐
│ 6.  Dependency Resolution                                   │
│    • Resolve requiredSkills dependencies                    │
│    • Sort by injectionOrder (0-100)                         │
│    • Detect circular dependencies                           │
└────────────────────┬────────────────────────────────────────┘
                     ↓
┌─────────────────────────────────────────────────────────────┐
│ 7.  Skill Injection & State Management                      │
│    • Read SKILL.md files                                    │
│    • Wrap in <skill> XML tags                               │
│    • Track loaded skills (prevent duplicates)               │
│    • Output enriched context to Claude                      │
└─────────────────────────────────────────────────────────────┘
```

### Key Concepts

**Intent Analysis**: Claude Haiku 4.5 analyzes your prompt to determine what
you're trying to do, assigning confidence scores to each skill.

**Confidence Thresholds**:
- **>0.65** - Required (automatically injected)
- **0.50-0.65** - Suggested (shown as optional)
- **<0.50** - Ignored

**Skill Types**:
- **Domain skills**: Provide comprehensive guidance (e.g.,
  `python-best-practices`)
- **Guardrail skills**: Enforce critical practices (e.g., `api-security`)

**Session Tracking**: Skills are tracked per conversation. Once loaded, they
won't be injected again in the same conversation.

**Caching**: Intent analysis results are cached for 1 hour using MD5 hashing of
(prompt + skills config).

## 🎓 Included Skills

### 1. skill-developer

A domain skill that guides creation and management of the skills system.

**Covers**: Skill structure, triggers, hooks, 500-line rule, progressive
disclosure

### 2. python-best-practices (Domain)

Python development best practices.

**Covers**: PEP 8, type hints, NumPy docstrings, common patterns, testing

### 3. git-workflow (Domain)

Git workflow and version control best practices.

**Covers**: Conventional Commits, branching, PRs, common operations, security

### 4. api-security (Guardrail)

API security and vulnerability prevention. Enforces security checks.

**Covers**: Authentication, input validation, SQL injection, XSS, OWASP Top 10

## 🏗️ Architecture

The system is built on three main components:

1. **Hook System** (`.claude/hooks/`)
   - `UserPromptSubmit` hook that analyzes prompts
   - 13 TypeScript modules handling the injection pipeline
   - Comprehensive test suite (11 test files)

2. **Skills** (`.claude/skills/`)
   - `SKILL.md` files with YAML frontmatter
   - Resource directories for detailed content
   - `skill-rules.json` configuration

3. **Configuration**
   - `skill-rules.json`: Defines all skills and triggers
   - `intent-analysis-prompt.txt`: AI analysis template
   - `.env`: API key configuration

For detailed architecture documentation, see
[docs/ARCHITECTURE.md](docs/ARCHITECTURE.md).

## 🛠️ Creating Custom Skills

Creating a new skill is simple:

### 1. Create Skill Directory

```bash
mkdir -p .claude/skills/my-skill/resources
```

### 2. Create SKILL.md

```markdown
---
name: my-skill
description: Brief description of what this skill does
---

# My Skill

## Purpose
What this skill helps with...

## When to Use This Skill
When it should activate...

[Your content here - keep under 500 lines]
```

### 3. Add to `skill-rules.json`

```json
{
  "my-skill": {
    "type": "domain",
    "autoInject": true,
    "requiredSkills": [],
    "description": "Brief description",
    "promptTriggers": {
      "keywords": ["keyword1", "keyword2"]
    }
  }
}
```

### 4. Test Activation

```bash
echo '{"session_id":"test","prompt":"mention keywords here"}' | \
  npx tsx .claude/hooks/skill-activation-prompt.ts
```

For comprehensive guidance, see
[docs/CREATING-SKILLS.md](docs/CREATING-SKILLS.md) or load the `skill-developer`
skill.

## 🔄 Keeping Skills Updated

As your codebase evolves, your skills need to stay in sync. Use the **`/wrap`
command** to maintain skill accuracy:

```bash
# When you finish a coding session
/wrap
```

**What `/wrap` does:**
- 🔍 **Analyzes** recent code changes and new patterns
- 📝 **Updates** skill documentation to reflect current codebase state
- 🎯 **Refines** keywords and triggers based on actual usage
- ✅ **Ensures** tests cover new functionality
- 📚 **Syncs** examples and best practices with implementation

**Workflow Example:**
1. Build a new feature (e.g., add authentication endpoints)
2. Run `/wrap` when done
3. The system updates `api-security` skill with your auth patterns
4. Updates keywords to include your specific auth methods
5. Adds examples from your actual implementation

This **continuous maintenance loop** keeps skills relevant and accurate as your
codebase grows. Skills that accurately reflect your current code are more likely
to be triggered at the right time.

> 💡 **Pro Tip:** Run `/wrap` at the end of each feature or sprint to keep
> skills synchronized with your evolving codebase.

## ⚙️ Configuration

### Change AI Model

Set the model used for intent analysis via environment variable (defaults to
`claude-haiku-4-5`):

```bash
# In .claude/hooks/.env
CLAUDE_SKILLS_MODEL=claude-sonnet-4-5  # Use Sonnet for more accurate analysis
```

Available models:

- `claude-haiku-4-5` (recommended: fast, cheap, accurate)
- `claude-sonnet-4-5` (more capable, higher cost)
- `claude-opus-4` (most capable, highest cost)

### Adjust Confidence Thresholds

All thresholds can be configured via environment variables (set in `.claude/hooks/.env`):

```bash
# In .claude/hooks/.env
SKILL_CONFIDENCE_THRESHOLD=0.65  # Auto-inject threshold (default: 0.65)
SKILL_SUGGESTED_THRESHOLD=0.50   # Suggested threshold (default: 0.50)
SKILL_SHORT_PROMPT_WORDS=6       # Skip AI for prompts under N words (default: 6)
SKILL_CACHE_TTL_MS=3600000       # Cache TTL in ms (default: 1 hour)
```

Or edit defaults in `.claude/hooks/lib/constants.ts`:

```typescript
export const CONFIDENCE_THRESHOLD = 0.65; // Auto-inject
export const SUGGESTED_THRESHOLD = 0.50; // Suggest
export const MAX_REQUIRED_SKILLS = 2; // Max required skills per prompt
export const MAX_SUGGESTED_SKILLS = 2; // Max suggested skills per prompt
```

### Adjust Cache TTL

Via environment variable (recommended):

```bash
SKILL_CACHE_TTL_MS=3600000  # 1 hour in milliseconds
```

Or edit `.claude/hooks/lib/constants.ts`:

```typescript
export const CACHE_TTL_MS = 60 * 60 * 1000; // 1 hour
```

### Enable Debug Mode

```bash
export CLAUDE_SKILLS_DEBUG=1
tail -f .claude/hooks/skill-injection-debug.log
```

## 🧪 Testing

Run the comprehensive test suite:

```bash
cd .claude/hooks
npm test
```

Test coverage includes:

- Intent analysis with caching
- Skill filtration and affinity injection
- Dependency resolution
- Session state management
- Output formatting
- Schema validation

## 🔍 Troubleshooting

### Skills not activating

```bash
# Check configuration
cat .claude/skills/skill-rules.json

# Test manually
echo '{"session_id":"test","prompt":"python"}' | \
  npx tsx .claude/hooks/skill-activation-prompt.ts

# Enable debug mode
export CLAUDE_SKILLS_DEBUG=1

# Check debug log
tail -f .claude/hooks/skill-injection-debug.log
```

### API key issues

```bash
# Verify .env file
cat .claude/hooks/.env

# Test API key
curl -H "x-api-key: $ANTHROPIC_API_KEY" \
  https://api.anthropic.com/v1/messages
```

### Clear caches

```bash
# Clear intent analysis cache
rm -rf .cache/intent-analysis/

# Clear session state
rm -rf .claude/hooks/state/
```

## 📚 Documentation

- [Getting Started Guide](docs/GETTING-STARTED.md) - Step-by-step setup
- [Architecture Documentation](docs/ARCHITECTURE.md) - System design and
  internals
- [Creating Skills Guide](docs/CREATING-SKILLS.md) - How to author custom skills
- [Hooks README](.claude/hooks/README.md) - Hook system documentation
- [Skills README](.claude/skills/README.md) - Skills system overview

## 🤝 Contributing

Contributions welcome! This is a template/framework for building skills systems.
Feel free to:

- Add new example skills
- Improve documentation
- Enhance the intent analysis prompt
- Add new features to the hook system
- Report bugs or suggest improvements

## 📄 License

MIT License - See [LICENSE](LICENSE) for details.

## 🙏 Credits

Built for the Claude Code community. Inspired by the need for automatic,
intelligent context loading in AI-assisted development workflows.

Special thanks to [@diet103](https://github.com/diet103) for the base reference
architecture.

## ❓ FAQ

**Q: Does this work with other AI assistants?**

**A**: The skills system is specific to Claude Code's hook system. However, the
concept can be adapted to other tools.

-----

**Q: How much does the API usage cost?**

**A**: Approximately $1-2/month at 100 prompts/day using Claude Haiku 4.5.
Caching significantly reduces costs.

-----

**Q: Can I use this without the AI analysis?**

**A**: Yes! Set `ANTHROPIC_API_KEY` to empty and the system will fall back to
keyword matching.

-----

**Q: How do I disable a skill?**

**A**: Set `autoInject: false` in `skill-rules.json` or remove the skill
entirely.

-----

**Q: What's the 500-line rule?**

**A**: Skills should stay under 500 lines for optimal agent processing. Use
progressive disclosure (resource files) for detailed content.

-----

**Q: Can skills depend on other skills?**

**A**: Yes! Use the `requiredSkills` array in `skill-rules.json` to specify
dependencies.

**Q: What's the difference between `requiredSkills` and `affinity`?**

| Aspect | `requiredSkills` | `affinity` |
|--------|---------------|----------|
| **Purpose** | Hard dependencies | Complementary skills |
| **Direction** | One-way only | Bidirectional (both ways) |
| **Slot cost** | Counts toward 2-skill limit | **FREE** - bonus injections |
| **Use when** | "Skill A needs B to work" | "Skills work great together" |

- Use **`requiredSkills`** for mandatory dependencies (e.g., advanced skill
  requires foundation)
- Use **`affinity`** for related skills that complement each other (loads 3+
  skills while using 1-2 slots)

-----

**Q: How do I see which skills are loaded?**

**A**: Look for the "📚 AUTO-LOADED SKILLS" banner at the start of Claude's
response.

claude-skills-supercharged

claude-skills-supercharged 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

claude-skills-supercharged

claude-skills-supercharged 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README