ai-skill-maker
ai-skill-maker is a meta-skill for designing, rendering, validating, and refreshing reusable AI skills and assistant instruction bundles. From rough intent, a repeatable workflow, or an existing repository, it generates a structured skill folder with evidence labels, validation scripts, release gates, and refresh-safe user rule blocks.
The goal is not a long prompt. It is a small, sharp capability that can be reused, tested, refreshed, and trusted.
Links
- Repository:
sscodeai/ai-skill-maker - License: Apache-2.0
- Runtime: Node.js 18+
- Form: Codex / agent personal skill (runtime payload is
SKILL.md,agents/,references/,assets/,scripts/)
Why it exists
A reusable AI skill is easy to draft and hard to keep trustworthy. A good skill needs narrow, accurate triggers; clear boundaries; an executable workflow; verifiable output; and a structure that future agents can refresh without erasing maintainer intent. ai-skill-maker provides that structure.
What you can build
| Output | Best for | Primary renderer |
|---|---|---|
| 🛠️ Functional skill | reports, PDFs, spreadsheets, API tasks, browser workflows, data processing | scripts/render-skill.mjs |
| 📄 Document/template skill | artifact structure, style, reusable assets, rendered fidelity | scripts/render-skill.mjs |
| 🔁 Workflow automation skill | repeatable procedures over tools, CLIs, APIs, scripts | scripts/render-skill.mjs |
| 🧭 Project maintainer skill | long-term AI maintainer guidance for a specific repository | scripts/render-project-skill.mjs |
| 🔌 Adapter instruction bundle | AGENTS.md, CLAUDE.md, Cursor rules, Copilot instructions |
scripts/render-adapter.mjs |
Design pillars
| Pillar | Meaning |
|---|---|
| 🎯 Root problem first | Design from the recurring problem, not the user’s first phrasing of it |
| 🧾 Evidence discipline | Separate claims into observed_fact, declared_intent, recommended_standard, inferred_assumption |
| 🛡️ Trust Gate | Explicit PASS/BLOCK review of permissions, sensitive data, dependencies, environment, external actions, rollback |
| 🧪 Evaluation skeleton | Ship trigger tests, output assertions, and BLOCK/ALLOW release gates that must be filled with real evidence |
| 🔒 Refresh safety | Preserve user-authored rule blocks and protected core blocks |
| 📏 File budget | Automatically check active Markdown instruction files against the 9,000-token ceiling |
Quick start
Requires Node.js 18+.
git clone https://github.com/sscodeai/ai-skill-maker.git
cd ai-skill-maker
node scripts/self-check.mjsCreate and validate a functional skill:
node scripts/render-skill.mjs --init-config functional > config.json
node scripts/validate-skill-config.mjs --input config.json --mode functional --strict
node scripts/render-skill.mjs --input config.json --output ./generated-skill --mode functional --strict
node scripts/validate-skill-output.mjs ./generated-skill
node scripts/file-budget.mjs ./generated-skillFor a project maintainer skill, run draft-project-config.mjs → validate-config.mjs --mode repo → render-project-skill.mjs → validate-project-skill.mjs.
Install as a Codex skill
node scripts/install-local-skill.mjs --dry-run # preview target and replacement behavior, no writes
node scripts/install-local-skill.mjs # install or syncWorkflow
| Step | Gate | Output |
|---|---|---|
| 1 | Identify the smallest useful capability | One-sentence root problem |
| 2 | Check permissions, secrets, dependencies, environment, external actions, rollback | Trust Gate PASS/BLOCK |
| 3 | Separate facts, intent, and assumptions | Evidence-labeled config |
| 4 | Choose mode and adapter | Renderer choice |
| 5 | Validate the config before rendering | Strict config check |
| 6 | Render or refresh | Skill folder or instruction bundle |
| 7 | Validate the output | structure, metadata, evidence labels, ledgers, user-rule markers, file budget |
| 8 | Release only with recorded evidence | Trigger tests, output assertions, license attribution |
Validation and release
node scripts/self-check.mjs
node scripts/check-core-principles.mjs
node scripts/file-budget.mjs
git diff --checkGenerated skills include references/evals/release-gate.md, and gates are BLOCK/ALLOW only. The validators cover required files, YAML metadata, evidence labels, file budget, and user-rule preservation markers — they do not prove a skill is absolutely safe, complete, or publish-ready.
Safety model
Generated skills should require explicit user permission before destructive operations, credential-sensitive work, remote publishing, force-pushing, visibility changes, and other high-risk actions. The maker’s own protected core principles live in references/rules/protected-core-principles.md and are locked by references/rules/core-principles.lock.json.
Next
- Read
SKILL.mdand the renderer scripts in the repository - See project-specific generation in ai-project-skill-maker
- Browse other SSCodeAI projects in Projects