A CLAUDE.md is just a markdown file at the root of your repo. Copy the content below into your own project's CLAUDE.md to give your agent the same context.
npx versuz@latest install libukai-awesome-agent-skills --kind=claude-mdcurl -o CLAUDE.md https://raw.githubusercontent.com/libukai/awesome-agent-skills/HEAD/CLAUDE.md# awesome-agentskills Development Guide
## Project Overview
This repository contains a collection of Claude Code plugins and agent skills for various development tasks. The main components include:
- **tldraw-helper**: Plugin for creating diagrams and visualizations using tldraw Desktop
- **agent-skills-toolkit**: Tools for creating, testing, and optimizing agent skills
- **plugin-dev**: Plugin development utilities and templates
## MCP Servers
This project integrates the following MCP (Model Context Protocol) servers:
### Excalidraw MCP
**Configuration**: `.claude/mcp.json`
A remote MCP server for creating hand-drawn style diagrams with interactive editing capabilities.
- **URL**: `https://mcp.excalidraw.com/mcp`
- **Type**: HTTP (remote)
- **Features**:
- Real-time hand-drawn diagram creation
- Interactive fullscreen editing
- Smooth viewport camera control
- Supports architecture diagrams, flowcharts, and creative visualizations
**Configuration Format**:
```json
{
"mcpServers": {
"excalidraw": {
"type": "http",
"url": "https://mcp.excalidraw.com/mcp"
}
}
}
```
**Important Notes**:
- MCP servers **cannot** be configured in global `~/.claude/settings.json`
- Use project-level `.claude/mcp.json` for project-specific MCP servers
- Restart Claude Code after adding new MCP servers
- Check MCP server approval status with `/mcp` command
**Usage**: The Excalidraw MCP tools are automatically available in Claude Code after restart. Use them for creating casual, hand-drawn style diagrams.
**Comparison with tldraw-helper**:
- **Excalidraw**: Hand-drawn, casual style, browser-based interaction
- **tldraw-helper**: Professional, precise style, tldraw Desktop integration
## Development Workflow
### Important: Local Development First
**All plugin and skill development should be done in this project directory first**, not in the marketplace cache (`~/.claude/plugins/cache/`).
The marketplace cache at `~/.claude/plugins/cache/awesome-agent-skills/` is where Claude Code loads plugins from when they're installed. However, this is a **read-only copy** for runtime use. Any modifications made there will be:
- Lost when the plugin is updated
- Not tracked in version control
- Not shared with other developers
### Correct Development Process
1. **Make changes in the project directory:**
```
/Users/likai/Github/Tools/awesome-agentskills/plugins/tldraw-helper/
```
2. **Test your changes locally** by symlinking or copying to the marketplace cache if needed
3. **Commit changes to git** in this project directory
4. **Publish updates** to the marketplace when ready
### Plugin Structure
Each plugin follows this structure:
```
plugins/plugin-name/
├── .claude-plugin/
│ └── plugin.json # Plugin manifest
├── commands/ # Slash commands
├── skills/ # Agent skills
├── agents/ # Subagents
├── hooks/ # Event hooks
├── README.md # Plugin documentation
└── CHANGELOG.md # Version history
```
## tldraw-helper Plugin
### Recent Improvements (2026-03-03)
**Critical Enhancement: Screenshot Verification**
Added mandatory screenshot verification step to prevent broken diagrams:
1. **Core Workflow** - Step 5 is now marked as "🚨 VERIFY VISUALLY (MANDATORY)"
2. **Detailed Checklist** - Added specific items to check in screenshots:
- Text readability and overlap
- Arrow positioning
- Layout cleanliness
- Element visibility
- Professional appearance
3. **Common Mistakes** - "Not analyzing screenshots" is now the #1 most critical error
4. **Key Takeaways** - Emphasized that API success ≠ good diagram
**Why This Matters:**
The tldraw API returns success even if the diagram is completely unusable (overlapping text, arrows crossing labels, elements off-screen). Visual verification is the ONLY way to know if a diagram is correct.
### Testing Changes
When testing tldraw-helper changes:
1. Modify files in `plugins/tldraw-helper/`
2. Copy to marketplace cache if needed:
```bash
cp -r plugins/tldraw-helper/* ~/.claude/plugins/cache/awesome-agent-skills/tldraw-helper/1.1.0/
```
3. Test with `/tldraw-helper:draw` command
4. Verify screenshot analysis is working correctly
## Git Workflow
### Current Status
```
M plugins/tldraw-helper/.claude-plugin/plugin.json
M plugins/tldraw-helper/README.md
M plugins/tldraw-helper/commands/draw.md
M plugins/tldraw-helper/skills/tldraw-canvas-api/SKILL.md
? plugins/tldraw-helper/CHANGELOG.md
? plugins/tldraw-helper/OPTIMIZATION_SUMMARY.md
? plugins/tldraw-helper/skills/tldraw-canvas-api/references/advanced-actions.md
```
### Committing Changes
When ready to commit:
```bash
# Stage specific files
git add plugins/tldraw-helper/skills/tldraw-canvas-api/SKILL.md
git add plugins/tldraw-helper/.claude-plugin/plugin.json
git add plugins/tldraw-helper/CHANGELOG.md
# Commit with descriptive message
git commit -m "Add mandatory screenshot verification to tldraw-helper skill
- Mark visual verification as mandatory step in workflow
- Add detailed checklist for screenshot analysis
- Promote 'not analyzing screenshots' to #1 critical error
- Emphasize that API success does not guarantee good diagrams
- Version bump to 1.1.0"
# Push to remote
git push origin main
```
## Best Practices
### Plugin Development
1. **Always update CHANGELOG.md** when making changes
2. **Bump version numbers** in plugin.json for significant changes
3. **Test thoroughly** before committing
4. **Document breaking changes** clearly
5. **Keep README.md up to date** with new features
### Skill Development
1. **Use clear, specific descriptions** for better triggering
2. **Include examples** in skill documentation
3. **Add error handling guidance** for common issues
4. **Test with real user scenarios**
5. **Iterate based on feedback**
### Agent Development
1. **Define clear triggering conditions** in agent descriptions
2. **Specify required tools** explicitly
3. **Provide system prompts** that guide behavior
4. **Test autonomous operation** thoroughly
5. **Document expected inputs/outputs**
## Publishing Updates
When ready to publish to the marketplace:
1. Ensure all changes are committed in this project
2. Update version numbers in plugin.json
3. Update CHANGELOG.md with release notes
4. Test the plugin thoroughly
5. Follow the marketplace publishing process
## Questions or Issues?
- Check existing documentation in plugin README files
- Review CHANGELOG.md for recent changes
- Test changes locally before committing
- Ask for clarification if workflow is unclear
---
**Remember: Always develop in this project directory, not in the marketplace cache!**