AI Skill Maker

日本語版

A meta-skill that builds, verifies, and governs other AI skills.

AI Skill Maker turns a rough idea, workflow, or repository into a reusable, runnable AI skill — then protects that skill from regressing as it evolves.

It fuses the strongest ideas from three open meta-skill projects:

Source	What we took	How it lands
CheshireMew/meta-skills	Behavior constitution	A protected core of 11 principles with a SHA-256 fingerprint lock; the maker refuses to silently weaken a skill's guarantees
yaojingang/yao-meta-skill	Evaluation & release gates	Every generated skill ships an evals/ skeleton: trigger tests, output assertions, and a BLOCK/ALLOW release gate
gnipbao/dao-skill	Root-problem thinking & Trust Gate	The maker settles the root problem before designing, and a hard Trust Gate blocks unsafe permissions, data leakage, or opaque dependencies

Why a meta-skill?

A skill is a set of instructions an AI agent reuses for a kind of work. Writing one well is hard: too vague and it misfires, too bloated and it wastes context, unverified and it silently degrades. AI Skill Maker is the skill for writing skills — with built-in checks that the result is actually runnable, verifiable, and safe.

What it creates

• Functional skills — reusable capabilities: presentations, Markdown reports, PDFs, spreadsheets, API workflows, browser automation
• Document/template skills — artifact structure, style, and rendering fidelity
• Workflow automation skills — deterministic repeated procedures via scripts and tools
• Project maintainer skills — long-lived guidance for a specific software repository
• Adapter instruction bundles — AGENTS.md, CLAUDE.md, Cursor rules, Copilot instructions

How it works

root problem → trust gate → evidence & intent → config → render → validate → release gate

1. Settle the root problem — the user's first phrasing is a symptom; the maker finds what they really need and the smallest capability that solves it
2. Run the Trust Gate — permissions, sensitive data, dependencies, environment, external actions, rollback. One BLOCK stops generation until the root cause is fixed
3. Gather evidence & intent — every claim is labeled observed_fact, declared_intent, recommended_standard, or inferred_assumption
4. Validate & render — JSON config → strict validation → templated skill folder
5. Ship with an evaluation skeleton — references/evals/ (trigger tests, output assertions, release gate) so the skill can prove it works in a fresh session
6. Protect the result — user-authored rule blocks and protected-core blocks are preserved on refresh; the constitution fingerprint lock fails loudly if the maker's own core drifts

Quick Start

Requires Node.js 18+.

# Install as a local Codex personal skill
node scripts/install-local-skill.mjs

# Generate a functional skill
node scripts/render-skill.mjs --init-config functional > config.json
node scripts/validate-skill-config.mjs --input config.json --strict
node scripts/render-skill.mjs --input config.json --output ./generated-skill --strict
node scripts/validate-skill-output.mjs ./generated-skill

The generated skill is a complete, standalone folder. A future agent session can use it without this maker.

Scripts

Script	Purpose
render-skill.mjs	Render a general skill (functional/document/workflow/refresh) from JSON config; --init-config drafts a starter
validate-skill-config.mjs	Strict-check a general skill config before rendering
validate-skill-output.mjs	Verify a rendered skill: required files, metadata, evidence labels, USER RULES blocks
render-project-skill.mjs	Render a project maintainer skill (genesis/repo mode)
validate-config.mjs	Strict-check a project maintainer config
validate-project-skill.mjs	Verify a rendered project maintainer skill
collect-repo-signals.mjs	Scan a repository into JSON signals for repo-mode drafting
draft-project-config.mjs	Draft a project maintainer config from repo signals
render-adapter.mjs	Render AGENTS.md / CLAUDE.md / Cursor / Copilot outputs, preserving user blocks
check-core-principles.mjs	Verify the maker's protected constitution fingerprint matches the lock
file-budget.mjs	Enforce the 9,000-token ceiling per active Markdown instruction file
self-check.mjs	Full repository health check (render, validate, budget, constitution, templates)
install-local-skill.mjs	Sync the runtime payload to ~/.codex/skills/ai-skill-maker

Repository layout

SKILL.md                  # the maker's own skill: routing, workflow, constitution pointer
agents/openai.yaml        # UI metadata
references/               # mode docs, checklists (intake, trust-gate, quality-gate), schemas, rules
assets/templates/skill/   # general skill template (SKILL.md + references incl. evals/)
assets/templates/project-skill/  # project maintainer template
scripts/                  # renderers, validators, checks, installer

Verification & Release

Run before committing or sharing:

node scripts/self-check.mjs          # full health check
node scripts/self-check.mjs --check-installed  # also verify the installed skill matches the repo
node scripts/check-core-principles.mjs         # constitution fingerprint
node scripts/file-budget.mjs                   # instruction-file budget

Generated skills carry their own release gate (references/evals/release-gate.md): trigger, output, structure, budget, trust, and license must all be ALLOW with recorded evidence before the skill is claimed release-ready. Missing evidence is a BLOCK, never an ALLOW.

License

Apache-2.0