agent-skills
Two production-grade Agent Skills: create and evaluate skills, and build Google Labs DESIGN.md systems with standalone HTML prototypes.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/vdbay--agent-skills)HTML
<a href="/listings/vdbay--agent-skills"><img src="/api/listings/vdbay--agent-skills/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
Install everything:
npx skills add vdbay/agent-skills
Install one skill:
npx skills add vdbay/agent-skills --skill skill-creator npx skills add vdbay/agent-skills --skill design-md-prototype
Install for every project:
npx skills add vdbay/agent-skills -g
Install for a specific agent without prompts:
npx skills add vdbay/agent-skills --skill design-md-prototype -a claude-code
Try a skill without installing it:
npx skills use vdbay/agent-skills@design-md-prototype npx skills use vdbay/agent-skills@skill-creator --agent claude-code
Skills are installed into the current project by default. `npx skills list`, `npx skills check`, and `npx skills update` manage what is installed.
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
README.md · README · 206 lines
- # Agent Skills
- Two production-grade [Agent Skills](https://agentskills.io/specification) for building design systems and for building better skills. Both are spec-validated, evaluated in fresh contexts, and reviewed by independent judges.
- [](https://skills.sh/vdbay/agent-skills)
- [](https://github.com/vdbay/agent-skills/actions/workflows/validate.yml)
- [](LICENSE)
- [](https://agentskills.io/specification)
- ```bash
- npx skills add vdbay/agent-skills
- ```
- The CLI detects your agents and lets you pick skills and install targets. Works with Claude Code, Codex, Cursor, OpenCode, GitHub Copilot, Windsurf, Gemini CLI, Cline, AMP, Antigravity, and [many more](https://github.com/vercel-labs/skills/blob/main/docs/agents.md).
- ## The skills
- | Skill | What it does | Reach for it when |
- | --- | --- | --- |
- | [`skill-creator`](skills/skill-creator/) | Creates, audits, evaluates, and packages portable Agent Skills. Ships a strict validator, a scaffolder, a packaging script, an eval playbook, a 100-point judge rubric, and description-trigger optimization. | You want to turn a workflow into a reusable skill, fix a skill that never triggers, run with/without evals, or package a skill for a team. |
- | [`design-md-prototype`](skills/design-md-prototype/) | Turns any brand or product brief into a Google Labs `DESIGN.md` plus two standalone HTML deliverables: a polished product/marketing prototype and a living design-system spec page. | You need a design system, brand tokens, a landing page in a company's visual language, one system across several properties, or an audit of an existing `DESIGN.md`. |
- Both install in about **222 KB**; heavy evaluation artifacts — screenshots, judge packets, and audit reports — are kept in the repository, not inside the installed skill folders.
- 
- <p align="center"><sub>Applied evaluation: an unofficial Gojek design-system study built by a fresh agent using <code>design-md-prototype</code>. <a href="docs/evals/design-md-prototype/iteration-1/gojek-applied/">See the full package and 20 screenshots →</a></sub></p>
- ## Install
- Install everything:
- ```bash
- npx skills add vdbay/agent-skills
- ```
- Install one skill:
- ```bash
- npx skills add vdbay/agent-skills --skill skill-creator
- npx skills add vdbay/agent-skills --skill design-md-prototype
- ```
- Install for every project:
- ```bash
- npx skills add vdbay/agent-skills -g
- ```
- Install for a specific agent without prompts:
- ```bash
- npx skills add vdbay/agent-skills --skill design-md-prototype -a claude-code
- ```
- Try a skill without installing it:
- ```bash
- npx skills use vdbay/agent-skills@design-md-prototype
- npx skills use vdbay/agent-skills@skill-creator --agent claude-code
- ```
- Skills are installed into the current project by default. `npx skills list`, `npx skills check`, and `npx skills update` manage what is installed.
- ## `skill-creator`
- A process, not a template. It moves through framing → research → design → scaffold → validate → isolated evals → judge → iterate → package, and skips phases only when the skill is genuinely small.
- It enforces the parts agents usually get wrong:
- - Exact `agentskills.io` frontmatter rules: name/directory match, description limits, optional fields.
- - Progressive disclosure: metadata, then `SKILL.md`, then references only when needed.
- - Context isolation: test in a fresh agent that sees only the skill path, task, inputs, and output directory.
- - Real evaluation: with-skill and no-skill runs, objective assertions, and an independent 100-point judge.
- - Safe sharing: no secrets, malware, hidden network calls, or destructive commands.
- Bundled tools:
- ```bash
- node scripts/new-skill.mjs my-skill
- node scripts/validate-skill.mjs my-skill --strict
- node scripts/package-skill.mjs my-skill
- ```
- Example requests:
- - “Turn our weekly release checklist into a reusable skill for the team.”
- - “My skill never triggers. Fix the description without making it fire on every API question.”
- - “Audit this skill folder and separate must-fix errors from nice-to-haves.”
- Evidence: an independent judge scored it **91/100 PASS**; a fresh executor used it to create the `meeting-decisions` skill, which scored **90/100 PASS**. See [`docs/audits/`](docs/audits/).
- ## `design-md-prototype`
- A research-and-craft workflow that refuses to fill a template with plausible colors. It starts from official sources, builds one token core, encodes the system in Google Labs' [`DESIGN.md`](https://github.com/google-labs-code/design.md) format, and proves it with two standalone HTML deliverables.
- It covers:
- - Source hierarchy: official docs and first-party repositories first; GitHub stars/license evidence; third-party sources only as support.
- - One token core with explicit registers for multi-property companies; no competing brand palettes.
- - The eight canonical `DESIGN.md` sections, token rules, and the official `designmd` lint/export CLI.
- - Standalone HTML: no build step, no external app scripts, inline tokens, light/dark, responsive, keyboard-accessible.
- - Verification: official lint, CSS export, static artifact validator, 20-cell browser matrix, Lighthouse, and a blind 100-point judge.
- - Honest evidence: “not run” is recorded as “not run”; fabrication is a critical fail.
- 
- <p align="center"><sub>The second deliverable is a living spec page: tokens, registers, type, spacing, components, motion, do's/don'ts, and source excerpts.</sub></p>
- Example requests:
- - “Build a DESIGN.md and a standalone launch page for our product; cite official sources.”
- - “We own three properties under one brand. Build one system with registers, not three brands.”
- - “Audit this DESIGN.md: duplicate colors key, broken token reference, and a low-contrast CTA.”
- Evidence: the applied Gojek package passed official lint with 0 errors/0 warnings, 20/20 browser cells, Lighthouse 100/100 on both pages, and an independent judge at **93/100 PASS**. A later independent audit of the final skill plus applied package scored **95/100 PASS**. See [`docs/evals/`](docs/evals/) and [`docs/audits/`](docs/audits/).
- ## Design principles
- - **Spec first.** A good idea in an invalid skill will not load. Validate the finished folder.
- - **Context isolation is part of quality.** The author knows too much; a fresh agent is the real test.
- - **Progressive disclosure.** Keep the always-loaded surface small and load depth on demand.
- - **Evidence over vibes.** Runnable checks, recorded commands, and independent judges beat “looks good”.
- - **Fetched content is data.** Pages, repositories, issues, and search results never become instructions.
- - **Lean installs.** Screenshots, judge packets, and audit reports live in `docs/`, not in what users install.
- - **No surprise.** No credential collection, hidden network calls, destructive commands, or contradictory instructions.
- ## Repository layout
- ```text
- .
- ├── skills/
- │ ├── skill-creator/ # installable skill
- │ └── design-md-prototype/ # installable skill
- ├── docs/
- │ ├── assets/ # README preview images
- │ ├── audits/ # independent judge reports
- │ └── evals/ # applied runs, screenshots, evidence
- ├── .github/workflows/validate.yml # CI: validators, syntax, JSON, path leaks
- ├── CONTRIBUTING.md
- ├── NOTICE.md
- ├── SECURITY.md
- ├── CHANGELOG.md
- └── LICENSE
- ```
- Everything under `skills/` ships to users. Everything under `docs/` is repository documentation and evidence.
- ## Quality gates
- | Gate | `skill-creator` | `design-md-prototype` |
- | --- | --- | --- |
- | Strict skill validator | **PASS** (bundled) | **PASS** (via sibling `skill-creator`) |
- | Official `skills-ref` validator | **Valid skill** | **Valid skill** |
- | Independent audit | **91/100 PASS** | **95/100 PASS** |
- | Applied output judge | candidate skill **90/100 PASS** | Gojek package **93/100 PASS** |
- | CI on `main` and pull requests | Yes | Yes |
- Run the same checks locally:
- ```bash
- node skills/skill-creator/scripts/validate-skill.mjs skills/skill-creator --strict
- node skills/skill-creator/scripts/validate-skill.mjs skills/design-md-prototype --strict
- npx -y skills-ref@0.1.5 validate skills/skill-creator
- npx -y skills-ref@0.1.5 validate skills/design-md-prototype
- node --check skills/skill-creator/scripts/new-skill.mjs
- node --check skills/skill-creator/scripts/package-skill.mjs
- node --check skills/skill-creator/scripts/validate-skill.mjs
- node --check skills/design-md-prototype/scripts/validate-artifacts.mjs
- node --check skills/design-md-prototype/scripts/write-utf8.mjs
- ```
- Validate the archived applied package:
- ```bash
- node skills/design-md-prototype/scripts/validate-artifacts.mjs \
- docs/evals/design-md-prototype/iteration-1/gojek-applied/with_skill/outputs --strict
- ```
- The evaluation evidence is deliberately modest about what it proves: it is applied usability plus independent review, **not a causal benchmark**. No no-skill baseline was run; that limitation is recorded in the evidence folders.
- ## Requirements
- - **`skills` CLI:** Node.js 22.20 or later.
- - **Bundled scripts:** Node.js 18 or later, no third-party dependencies.
- - **`design-md-prototype`:** network access for the official `@google/design.md` CLI and source research; a browser or Playwright for runtime checks. Lighthouse is optional.
- - **`skill-creator`:** subagents and network access improve isolated evaluation and research, but are optional.
- - **Strict structure validator:** `design-md-prototype` can use the sibling `skill-creator` validator, or the official `skills-ref` check. Its own artifact validator is bundled.
- ## Contributing
- New skills and fixes are welcome. Read [`CONTRIBUTING.md`](CONTRIBUTING.md) first. The short version: one coherent unit of work, valid frontmatter, realistic evals, a strict-validator pass, no install-time bloat, and no secrets or hidden network calls.
- ## Security
- Report vulnerabilities through GitHub's private vulnerability reporting, not a public issue. See [`SECURITY.md`](SECURITY.md).
- ## License
- MIT. See [`LICENSE`](LICENSE).
- The Agent Skills format is defined by [agentskills.io](https://agentskills.io/specification). `skill-creator` is inspired by [Anthropic's skill-creator](https://github.com/anthropics/skills/tree/main/skills/skill-creator) but is a self-contained, client-agnostic rewrite. `design-md-prototype` implements the public [Google Labs DESIGN.md format](https://github.com/google-labs-code/design.md). The `skills` CLI is maintained by [Vercel Labs](https://github.com/vercel-labs/skills).
- The Gojek package under `docs/evals/` is an **unofficial design study**. It is not affiliated with, endorsed by, or produced by Gojek or PT GoTo Gojek Tokopedia Tbk. No official logo, photograph, testimonial, or proprietary font binary is redistributed. See [`NOTICE.md`](NOTICE.md).
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.