Writing Skills

Writing Skills (Meta-Skill)

Adapted from obra/superpowers

Overview

A guide for creating effective, discoverable, and well-tested skills. Use this when authoring new skills for this repo.

What is a Skill?

A skill is a reference guide for proven techniques, patterns, or tools. Skills help AI agents find and apply effective approaches.

Skills are: Reusable techniques, patterns, tools, reference guides Skills are NOT: Narratives about how you solved a problem once

When to Create a Skill

Create when:

• Technique wasn't intuitively obvious
• You'd reference this again across projects
• Pattern applies broadly (not project-specific)
• Others would benefit

Don't create for:

• One-off solutions
• Standard practices well-documented elsewhere
• Project-specific conventions (put in CLAUDE.md instead)
• Mechanical constraints enforceable with linting/validation

Skill Types

• Technique: Concrete method with steps (condition-based-waiting, root-cause-tracing)
• Pattern: Way of thinking about problems (flatten-with-flags, test-invariants)
• Reference: API docs, syntax guides, tool documentation

SKILL.md Structure

Every skill is a skills/<skill-name>/SKILL.md file that opens with YAML frontmatter, then markdown:

---
name: skill-name
description: One line — lead with the trigger use case (this drives auto-invocation).
# disable-model-invocation: true   # command-only: won't auto-trigger; run as /skill-name
# paths: ["**/*.py"]               # auto-trigger only when working on matching files
---

# Skill Name

## Overview
What is this? Core principle in 1-2 sentences.

## When to Use
Bullet list with SYMPTOMS and use cases.
When NOT to use.

## Core Pattern
Before/after comparison or step-by-step process.

## Quick Reference
Table or bullets for scanning common operations.

## Common Mistakes
What goes wrong + fixes.

Discovery Optimization

Future Claude instances need to FIND your skill. Optimize for this:

Description: Start with "Use when..." to focus on triggering conditions. Describe the problem, not the solution. NEVER summarize the skill's workflow in the description — Claude may follow the description shortcut instead of reading the full skill.

# BAD: Summarizes workflow
description: Use when executing plans - dispatches subagent per task with code review

# GOOD: Just triggering conditions
description: Use when implementing any feature or bugfix, before writing implementation code

Keywords: Use words Claude would search for — error messages, symptoms, synonyms, tool names.

Naming: Kebab-case directory names. For a command-only skill the directory name is the slash command (scaffold, verify-django, nextjs-bootstrap) — name it exactly what you'd type. For an auto-triggering skill use a clear noun/verb phrase that matches its trigger (django-patterns, work-discipline, plan-and-scope). Match the repo's existing names; don't force gerunds.

Token Efficiency

Every token in frequently-loaded skills counts.

• Move details to supporting files, reference with cross-links
• Use cross-references instead of repeating content
• Compress examples — one excellent example beats many mediocre ones
• Don't repeat what's in cross-referenced skills
• Target: <200 words for frequently-loaded skills, <500 words for others

Code Examples

One excellent example beats many mediocre ones. Choose the most relevant language for the pattern. Make it complete, runnable, and well-commented. Don't implement in 5+ languages.

File Organization

skills/
  skill-name/
    SKILL.md              # Main reference (required)
    supporting-file.*     # Only if needed (heavy reference, reusable tools)

Keep content inline unless it exceeds ~100 lines of reference material or is a reusable script/utility.

Testing Skills

Before deploying a skill, validate it works:

1. Baseline: Run a scenario WITHOUT the skill — document what the agent does wrong
2. With skill: Run the same scenario WITH the skill — verify the agent now complies
3. Edge cases: Try pressure scenarios where the agent might rationalize skipping the skill
4. Close loopholes: If the agent finds workarounds, add explicit counters

Different skill types need different test approaches:

• Discipline skills (rules/requirements): Test with pressure scenarios
• Technique skills (how-to): Test with application scenarios and edge cases
• Pattern skills (mental models): Test recognition and counter-examples
• Reference skills (docs): Test retrieval and correct application

Anti-Patterns

• Narrative style: "In session 2025-10-03, we found..." — too specific, not reusable
• Multi-language dilution: examples in 5 languages — mediocre quality, maintenance burden
• Code in flowcharts: Can't copy-paste, hard to read
• Generic labels: "step1", "helper2" — labels should have semantic meaning
• Untested skills: If you didn't watch an agent fail without the skill, you don't know if it teaches the right thing

Checklist for New Skills

• Name uses letters, numbers, hyphens only
• Clear overview with core principle in 1-2 sentences
• "When to Use" section with specific triggers/symptoms
• Keywords throughout for search discovery
• One excellent code example (if applicable)
• Quick reference table for scanning
• Common mistakes section
• Tested against baseline behavior
• No narrative storytelling
• Supporting files only for heavy reference or reusable tools