Skill Creator

When to Create a Skill

Create a skill when:

• A pattern is used repeatedly and AI needs guidance
• Project-specific conventions differ from generic best practices
• Complex workflows need step-by-step instructions
• Decision trees help AI choose the right approach

Don't create a skill when:

• Documentation already exists (create a reference instead)
• Pattern is trivial or self-explanatory
• It's a one-off task

Skill Structure

Choose {skill-root} as domains/{domain}/skills for one owner or skills for multiple owners.

{skill-root}/{skill-name}/
├── SKILL.md              # Required - main skill file
├── assets/               # Optional - generated templates, schemas, fixtures, examples
│   ├── template.py
│   └── schema.json
└── references/           # Optional - longer guidance, edge cases, explanatory docs
    └── guidance.md

SKILL.md Template

---
name: {skill-name}
description: >
  {One-line description of what this skill does}.
  Trigger: {When the AI should load this skill}.
license: Apache-2.0
metadata:
  author: {author}
  version: "1.0.0"
  status: backlog
---

## When to Use

{Bullet points of when to use this skill}

## Critical Patterns

{The most important rules - what AI MUST know}

## Code Examples

{Minimal, focused examples}

## Commands

```bash
{Common commands}

Resources

• Templates: See assets/ for {description}
• Documentation: See references/ for local docs

---

## Naming Conventions

| Type | Pattern | Examples |
|------|---------|----------|
| Generic skill | `{owner}/skills/{technology}` | `domains/plan/skills/java-testing` |
| Project-specific | `{owner}/skills/{project}-{component}` | `domains/meta/skills/myapp-api` |
| Testing skill | `{owner}/skills/{project}-test-{component}` | `domains/plan/skills/myapp-test-api` |
| Workflow skill | `{owner}/skills/{action}-{target}` | `domains/meta/skills/skill-creator` |

---

## Decision: assets/ vs references/

Need generated templates? → assets/ Need JSON schemas? → assets/ Need fixtures/examples? → assets/ Need conceptual guidance? → references/ Need edge cases? → references/

**Key Rule**: concrete generated material goes in `assets/`; explanatory material goes in `references/`.

---

## Frontmatter Fields

| Field | Required | Description |
|-------|----------|-------------|
| `name` | Yes | Skill identifier (lowercase, hyphens) |
| `description` | Yes | What + Trigger in one block |
| `license` | Yes | Skill license, commonly `MIT` or `Apache-2.0` |
| `metadata.author` | Yes | Skill author or maintainer |
| `metadata.version` | Yes | Semantic version as string |
| `metadata.status` | Yes | `backlog`, `in-progress`, `testing`, or `done` |

---

## Content Guidelines

### DO
- Start with the most critical patterns
- Use tables for decision trees
- Keep code examples minimal and focused
- Include Commands only when repeatable commands are part of the contract

### DON'T
- Add Keywords section (agent searches frontmatter, not body)
- Duplicate content from existing docs (reference instead)
- Include lengthy explanations (link to docs)
- Add troubleshooting sections (keep focused)
- Use web URLs in references (use local paths)

---

## Behavioral Evaluation (Conditional)

The manual-catalog linter (`python3 scripts/lint-manual-tests.py`) validates case metadata, links, and changed-runtime coverage, not model behavior. When a high-impact discipline skill shows a reproducible agent failure despite being available, run the bounded evaluation pilot in [references/behavioral-evaluation.md](references/behavioral-evaluation.md): pre-written criteria, a no-guidance control, multiple fresh-context runs per arm, preserved raw receipts, manual scoring, and reported variance and cost. It is never a routine step or a CI gate.

---

## Registering the Skill

After creating or changing a skill, keep `metadata.version` and `metadata.status` accurate. Use a patch bump for wording/path/template/internal contract fixes, a minor bump for new capabilities or optional flows, and a major bump for breaking activation/output behavior. Put a one-domain skill directly under that domain. For multiple domains, keep one body at `skills/<skill-name>` and link every owner with `domains/<domain>/skills/<skill-name> -> ../../../skills/<skill-name>`.

---

## Checklist Before Creating

- [ ] Skill doesn't already exist (check `domains/*/skills/*/SKILL.md` and `skills/*/SKILL.md`)
- [ ] One-domain body is `domains/{domain}/skills/{skill-name}/`
- [ ] Multi-domain body is `skills/{skill-name}/` with a symlink from every owner
- [ ] Frontmatter `name` equals the skill directory basename
- [ ] Pattern is reusable (not one-off)
- [ ] Name follows conventions
- [ ] Frontmatter is complete (description includes trigger keywords)
- [ ] `metadata.version` is strict SemVer
- [ ] `metadata.status` reflects the lifecycle state
- [ ] Critical patterns are clear
- [ ] Code examples are minimal
- [ ] Commands section exists only when useful
- [ ] Existing manual coverage key is updated, or one new immutable manual case is added

## Resources

- **Templates**: See [assets/](assets/) for SKILL.md template