Creating New Skills

When to Create a Skill

After completing work, ask yourself:

1. •Would this procedure help on similar future tasks?
2. •Did I make mistakes that a checklist would prevent?
3. •Is this complex enough that future-me might skip steps?
4. •Did I learn a pattern worth codifying?

If yes to any → Extract to skill.

Skill Structure Template

---
name: Clear Descriptive Name
when_to_use: Specific trigger conditions that indicate this skill applies
---

# Overview

Core principle in 1-2 sentences. What's the key insight?

# Implementation Checklist

- [ ] First step
- [ ] Second step
- [ ] Third step
- [ ] Verify results

# Quality Standards

Specific criteria for success:
- What must be true when done?
- What errors must be zero?
- What verification proves it works?

# Anti-Patterns

Common mistakes to avoid:
❌ Bad pattern - why it's wrong
✅ Good pattern - why it's right

# Supporting

Links to tools, scripts, or reference docs:
- Tool documentation: path/to/tool/docs
- Related skills: path/to/related/SKILL.md

Testing Process

Before adding a skill to the library:

1. •Write the skill following template above
2. •Self-test by following it yourself for a similar task
3. •Test on subagent (optional but recommended):
   • •Dispatch subagent with realistic task
   • •Give subagent access to the skill
   • •Watch for places subagent might skip steps or misinterpret
4. •Strengthen language where compliance is weak
5. •Retest until bulletproof

Adding to Library

# Create skill directory and file
mkdir -p ~/.config/agent/skills/skills/[category]/[skill-name]
vim ~/.config/agent/skills/skills/[category]/[skill-name]/SKILL.md

# Test by using it yourself

# Commit when satisfied
cd ~/.config/agent/skills
git add skills/[category]/[skill-name]/SKILL.md
git commit -m "Add [skill-name] skill for [purpose]"

# Push if you're working from a fork
git push origin main

Categories

Organize skills into logical categories:

• •meta/ - Skills about skills (like this one)
• •research/ - Information gathering and verification
• •data-analysis/ - Working with data, spreadsheets, visualization
• •communication/ - Writing, reporting, documentation
• •automation/ - Task automation, scripting, workflows
• •tools/ - Reference docs for specific tools (supporting only)

Skill vs Tool Documentation

Skill: Process-focused, tells you WHAT to do and WHEN

• •"When analyzing data, always verify with multiple sources"
• •"Use formulas instead of hardcoded values"
• •"Test edge cases before declaring complete"

Tool doc: Tool-focused, tells you HOW to use specific technology

• •"pandas.read_excel() loads Excel files"
• •"openpyxl preserves formulas"
• •"WebFetch retrieves web content"

Put tool docs in skills/tools/[tool-name]/ as supporting material. Put processes in skills/[category]/[skill-name]/.

Quality Checklist for New Skills

Before committing a new skill:

• • Has clear name and when_to_use in frontmatter
• • Includes Overview section with core principle
• • Has Implementation checklist with actionable steps
• • Defines Quality Standards or success criteria
• • Includes Anti-Patterns section if relevant
• • Has been self-tested on realistic scenario
• • Language is strong enough to enforce compliance
• • Placed in appropriate category

Self-Improvement Loop

This is the key to the system:

Agent encounters task
→ Uses existing skill (or creates new pattern)
→ Learns something reusable
→ Writes new skill
→ Future agents use it
→ System gets smarter

Your job is to recognize when patterns are reusable and codify them. The library grows organically as you work.