Skill Authoring Guide
This guide defines the required structure and quality bar for skills in this repository. Follow it when creating a new skill or making substantial changes to an existing one.
What makes a good skill
A skill earns its place when it satisfies all of these:
- Distinct scope. It covers a job the existing skills do not, or covers a shared job for a meaningfully different audience/vehicle class (e.g., Motorcycle Mechanic vs. Mechanic Assistant).
- Workflow, not trivia. The skill teaches a way of working with the AI — context to gather, order of operations, iteration pattern — not just a topic list.
- Honest limitations. It states clearly what the AI cannot do in this domain and where professional help is mandatory.
- Safe by design. It cannot be followed into a hazard by a beginner taking it literally.
Directory layout
skills/
└── your-skill-name/ # kebab-case, descriptive
├── README.md # the skill definition (required core)
├── README.ja.md # Japanese version (required for merged skills)
├── SKILL.md # Claude Code native skill (see below)
└── skill.yaml # machine-readable metadata (see below)
Claude Code definition (SKILL.md)
Each skill folder also contains a SKILL.md so the repository works as a Claude Code plugin. Keep it thin — frontmatter (name matching the folder, description stating when to use the skill with concrete trigger phrases, license) plus the shared operating rules and a pointer to README.md, which remains the source of truth. Copy an existing skill’s SKILL.md and adjust the title and description; do not duplicate README content into it.
Machine-readable definitions (skill.yaml)
Each skill folder contains a skill.yaml for agent frameworks and tooling:
schema_version: 1 # bump only on breaking schema changes
name: your-skill-name # must match the folder name
title: "Your Skill Name"
summary: >-
One or two sentences; used for routing and skill registries.
audience: "Who this is for"
tags: [lowercase, kebab-case, searchable, terms]
languages: # relative paths to each language's entry file
en: README.md
ja: README.ja.md
related: [other-skill-names] # must reference existing skill folder names
safety_notes: "Verify all safety-critical specifications against factory documentation. See docs/safety-disclaimer.md."
The YAML carries metadata only — the README remains the source of truth for skill content. Keep summary and tags in sync when the README’s scope changes.
Required README structure
Use this template. Every section is mandatory, in this order.
# Skill Name
One-sentence summary of what the skill does and its scope.
## Description
2–4 sentences: the problem this skill solves, its approach, and how it
relates to (not replaces) professional resources.
**Audience:** who this is for, including experience range.
## Capabilities
- Bulleted list of concrete things the skill helps with
- Group with bold sub-headers if there are more than ~8 items
- Be specific: "interpret freeze frame data" beats "help with diagnostics"
## Example Prompts
At least 4 realistic prompts in fenced ```text blocks. Requirements:
- Use plausible, specific vehicles (real year/make/model/engine)
- Show the *variety* of the skill (diagnosis, planning, explanation…)
- Model good prompting: include the context the workflow says to gather
## Recommended Workflow
Numbered steps describing how to run a session with this skill:
1. What context to gather first
2. How to phrase the initial request
3. How to iterate (report findings, answer follow-ups)
4. How to verify results before acting
## Limitations
The most important section. Must cover, as applicable:
- What the AI cannot perceive or verify in this domain
- Where factory documentation overrides the assistant
- Which tasks must be deferred to professionals, and why
- Knowledge-cutoff and regional-variation caveats
## Related Skills
- Relative links to 2–3 skills this one hands off to or from,
with a phrase explaining the relationship
Writing rules
- Voice: professional, direct, practical. Write like a good shop foreman explains things — respectful of beginners, never condescending.
- Specificity: every capability and example should survive the question “could a reader act on this?”
- Specs: never present torque values, capacities, or clearances as authoritative. Frame as typical ranges + “verify against the factory manual.”
- Safety-critical systems (brakes, steering, SRS, fuel, high-voltage): mention them only with explicit professional-deferral language.
- Prohibited content: anything enabling emissions-control defeat, safety-system defeat, odometer fraud, or immobilizer/anti-theft bypass. No exceptions.
- Links: relative paths only, so they work on GitHub and in clones.
When you add a skill, also update
skills/README.md— add a row to the index table and, if appropriate, the “Which skill do I need?” listREADME.md— add a row to the Skills table (and remove it from Roadmap if it was listed)README.ja.md— add the corresponding row (English name + short Japanese gloss is fine)skill.yaml— create it per the schema above; add your skill to existing skills’relatedlists where the handoff is realREADME.ja.mdin the skill folder — a Japanese version is required before merge (maintainers can help translate; open the PR and say so)CHANGELOG.md— add an entry under [Unreleased]
Quality checklist
- Scope does not substantially duplicate an existing skill
- All seven sections present, in order
- ≥ 4 example prompts, each with realistic vehicle context
- Workflow is actionable step-by-step, not a restatement of capabilities
- Limitations are specific to this domain (not boilerplate)
- No authoritative safety-critical specs; professional deferrals present
- Index files and changelog updated
- All links resolve