writing-skills

Writing Skills (Excellence)

Compatibility: these templates are conventions, not a universal loader schema. Use the target host's documented skill directory and current schema; custom metadata.triggers is optional and does not itself guarantee discovery. Bundled upstream guidance is a historical snapshot. Rendering requires Node.js and a locally installed Graphviz dot; inspect trusted DOT only, and open generated SVG cautiously. --combine shares node IDs across graphs and is a visual convenience, not a semantics-preserving merge. Outputs are exclusive: remove or choose another output directory deliberately before rerunning.

Scope: examples are authoring/test fixtures, not live instructions. Follow the user, host, and repository instruction hierarchy. Run pressure tests only with authorized isolated fixtures; never delete user work or weaken safety/approval boundaries to enforce a skill. Preserve existing work and adapt the validation plan when a baseline already passes. Sample compliance does not prove universal reliability.

Dispatcher for skill creation excellence. Use the decision tree below to find the right template and standards.

⚡ Quick Decision Tree

What do you need to do?

1. Create a NEW skill:

   • Is it simple (single file, <200 lines)? → Tier 1 Architecture
   • Is it complex (multi-concept, 200-1000 lines)? → Tier 2 Architecture
   • Is it a massive platform (10+ products, AWS, Convex)? → Tier 3 Architecture

2. Improve an EXISTING skill:

   • Fix "it's too long" -> Modularize (Tier 3)
   • Fix "AI ignores rules" -> Anti-Rationalization
   • Fix "users can't find it" -> CSO (Search Optimization)

3. Verify Compliance:

   • Check metadata/naming -> Standards
   • Add tests -> Testing Guide

📚 Component Index

Component	Purpose
CSO	"SEO for LLMs". How to write descriptions that trigger.
Standards	File naming, YAML frontmatter, directory structure.
Anti-Rationalization	How to write rules that agents won't ignore.
Testing	How to ensure your skill actually works.

🛠️ Templates

• Technique Skill (How-to)
• Reference Skill (Docs)
• Discipline Skill (Rules)
• Pattern Skill (Design Patterns)

When to Use

• Creating a NEW skill from scratch
• Improving an EXISTING skill that agents ignore
• Debugging why a skill isn't being triggered
• Standardizing skills across a team

How It Works

1. Identify goal → Use decision tree above
2. Select template → From references/templates/
3. Apply CSO → Optimize description for discovery
4. Add anti-rationalization → For discipline skills
5. Test → RED-GREEN-REFACTOR cycle

Quick Example

---
name: my-technique
description: Use when [specific symptom occurs].
metadata:
  category: technique
  triggers: error-text, symptom, tool-name
---

# My Technique

## When to Use
- [Symptom A]
- [Error message]

Common Mistakes

Mistake	Fix
Description summarizes workflow	Describe the capability and specific use conditions concisely
No metadata.triggers	Optional host-specific metadata; include useful terms in description
Generic name ("helper")	Use a descriptive lowercase hyphenated ID
Long monolithic SKILL.md	Split into references/

See gotchas.md for more.

✅ Pre-Deploy Checklist

Before deploying any skill:

• name field matches directory name exactly
• SKILL.md filename is ALL CAPS
• Description explains what the skill does and when to use it
• Frontmatter matches the target host and repository schema
• Total lines < 500 (use references/ for more)
• No @ force-loading in cross-references
• Tested with real scenarios

🔗 Related Skills

• opencode-expert: For OpenCode environment configuration
• If your host defines /write-skill, it may provide guided creation; that command is not bundled here

Examples

Create a Tier 1 skill:

mkdir -p ~/.config/opencode/skills/my-technique
touch ~/.config/opencode/skills/my-technique/SKILL.md

Create a Tier 2 skill:

mkdir -p ~/.config/opencode/skills/my-skill/references/core
touch ~/.config/opencode/skills/my-skill/{SKILL.md,gotchas.md}
touch ~/.config/opencode/skills/my-skill/references/core/README.md

Limitations

• Use this skill only when the task clearly matches the scope described above.
• Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.