Skill Creator

Skill Creator

Create effective skills using TDD methodology with scaffolding scripts.

Overview

Skills are TDD applied to process documentation. Write test cases (pressure scenarios), watch them fail (baseline behavior), write the skill, watch tests pass, and refactor.

Core principle: If you didn't watch an agent fail without the skill, you don't know if the skill teaches the right thing.

Quick Start

# Initialize new skill
scripts/init_skill.py my-skill --path ~/.claude/skills

# Validate skill structure
scripts/quick_validate.py ~/.claude/skills/my-skill

# Package for distribution
scripts/package_skill.py ~/.claude/skills/my-skill

Skill Structure

skill-name/
├── SKILL.md              # Main reference (required)
├── scripts/              # Executable code (optional)
├── references/           # Documentation loaded as needed (optional)
└── assets/               # Templates, images, fonts (optional)

SKILL.md Frontmatter

Only name and description fields. Max 1024 chars total.

---
name: hyphen-case-name
description: Use when [specific triggering conditions]. Start with "Use when..."
---

CRITICAL: Description = when to use, NOT what skill does. Testing revealed that workflow summaries in descriptions cause agents to skip reading the full skill.

# ❌ BAD: Summarizes workflow
description: Use when executing plans - dispatches subagent with code review between tasks

# ✅ GOOD: Just triggers
description: Use when executing implementation plans with independent tasks

SKILL.md Body Structure

# Skill Name

## Overview
Core principle in 1-2 sentences.

## When to Use
Symptoms and use cases (bullets). When NOT to use.

## Core Pattern
Before/after or key workflow.

## Quick Reference
Table or bullets for scanning.

## Common Mistakes
What goes wrong + fixes.

Creation Process (TDD)

1. Understand with Examples

Ask: "What functionality should this skill support?" "Can you give usage examples?"

2. RED: Baseline Test

Run pressure scenario WITHOUT skill. Document:

• What choices did agent make?
• What rationalizations (verbatim)?
• Which pressures triggered violations?

3. GREEN: Write Minimal Skill

Address those specific rationalizations. Don't add hypothetical content.

Run scenarios WITH skill. Agent should comply.

4. REFACTOR: Close Loopholes

New rationalization found? Add explicit counter. Re-test until bulletproof.

See references/tdd-testing.md for full testing methodology.

Token Efficiency

Target word counts:

• Frequently-loaded skills: <200 words
• Other skills: <500 words

Techniques:

• Move details to --help in scripts
• Cross-reference other skills instead of repeating
• Compress examples

Progressive Disclosure

1. Metadata - Always in context (~100 words)
2. SKILL.md body - When skill triggers (<5k words)
3. Bundled resources - As needed (unlimited)

For skills with variants (frameworks, providers), keep core in SKILL.md, move variant details to references/:

cloud-deploy/
├── SKILL.md (workflow + selection)
└── references/
    ├── aws.md
    ├── gcp.md
    └── azure.md

Bundled Resources

scripts/

Deterministic code for fragile or repeated operations. May execute without loading into context.

references/

Documentation loaded on-demand. For 100+ line files, include table of contents.

assets/

Templates, images, fonts used in output (not loaded into context).

CSO (Claude Search Optimization)

Rich description: Include concrete triggers, symptoms, situations. Write in third person.

Keywords: Error messages, symptoms, synonyms, tool names.

Naming: Verb-first, hyphen-case: creating-skills not skill-creation.

Anti-Patterns

| Anti-Pattern | Why Bad | |--------------|---------| | Narrative example ("In session 2025-10-03...") | Too specific, not reusable | | Multi-language examples | Mediocre quality, maintenance burden | | Code in flowcharts | Can't copy-paste | | Generic labels (step1, helper2) | Labels need semantic meaning |

Flowchart Usage

Use ONLY for:

• Non-obvious decision points
• Process loops where you might stop early
• "When to use A vs B" decisions

Never for: Reference material, code examples, linear instructions.

The Iron Law

NO SKILL WITHOUT A FAILING TEST FIRST

This applies to NEW skills AND EDITS. Write skill before testing? Delete it. Start over.

Checklist

RED Phase:

• [ ] Create pressure scenarios
• [ ] Run WITHOUT skill, document baseline verbatim
• [ ] Identify rationalization patterns

GREEN Phase:

• [ ] Name: hyphen-case, letters/numbers/hyphens only
• [ ] Description: starts with "Use when...", no workflow summary
• [ ] Address specific baseline failures
• [ ] Run WITH skill, verify compliance

REFACTOR Phase:

• [ ] Identify NEW rationalizations
• [ ] Add explicit counters
• [ ] Re-test until bulletproof

Quality:

• [ ] Flowchart only if decision non-obvious
• [ ] Quick reference table
• [ ] Common mistakes section
• [ ] No narrative storytelling

Deploy:

• [ ] Run quick_validate.py
• [ ] Package with package_skill.py (optional)

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: letitbk
• Source: letitbk/claude-academic-setup
• License: MIT

Install and usage instructions live in the source repository linked above.