Hairyf Skills - Setup Guide¶

Source: hairyf/skills via skills.sh - 25.7K combined installs across 80 indexed listings; first documented 2026-11-03 (nightly sweep)

GitHub: hairyf/skills - 26 stars, MIT (LICENSE.md, copyright Hairyf), branch main, last pushed 2026-10-14; 80 skill directories under skills/, one SKILL.md each plus a references/ folder of topic files, 1,966 Markdown files and about 11 MB in total, with upstream documentation cloned as git submodules under sources/

Category: Framework and tooling knowledge for agents across Vue, Nuxt, Vite, React, testing and bundling, plus scaffolding starters and a small media and Minecraft wing

Quality Tier: 🟢 Green - all nine sampled verdicts across three listings pass on Gen Agent Trust Hub, Socket and Snyk, and the repository ships an MIT licence file

Hairy's Skills is a fork of antfu/skills, maintained by Hairyf, and the README is honest about what it is: a curated collection of Agent Skills built on Anthony Fu's collection and extended with skills reflecting Hairyf's own preferences and projects. It doubles as a proof of concept for generating skills from source documentation and keeping them in sync. meta.ts maps upstream repositories such as react.dev, vuejs/docs, nuxt/nuxt, tailwindcss.com and motiondivision/motion to output skill names, .gitmodules clones those repositories into sources/, and the generated output is committed under skills/.

The tree holds 80 skill directories and 1,966 Markdown files. Coverage runs from Vue, Nuxt, Vite, UnoCSS, Pinia and Vitest through React and Next.js to Tailwind CSS, Motion and Anime.js, then out to Tauri, Electron, Flutter and React Native, OpenAPI 2.0 and 3.2, and a set of arch-* starter templates. A handful have nothing to do with front-end work: imagine, sonic and vision for media generation, and a Minecraft group covering mods, models, textures and shader work. The README's own warning belongs up front rather than in a footnote: "I haven't fully tested how well the skills perform in practice".

Installation¶

The README gives two commands, both through the skills CLI:

# the whole collection
pnpx skills add hairyf/skills

# the whole collection, installed globally
pnpx skills add hairyf/skills --all -g

Install one skill at a time instead when you are wiring this into a working profile. Eighty skills mean eighty trigger descriptions competing for the same requests:

npx skills add hairyf/skills --skill nuxt
npx skills add hairyf/skills --skill vue-best-practices

The manual route is a clone plus a copy, because every skill is self-contained under skills/:

git clone --depth 1 https://github.com/hairyf/skills.git /tmp/hairyf-skills
ls -d /tmp/hairyf-skills/skills/*/ | wc -l
cp -r /tmp/hairyf-skills/skills/nuxt ~/.hermes/profiles/corpusiq/skills/

Copy the whole directory, not only SKILL.md. The generated skills are index pages: the nuxt skill opens with a topic table and routes each row to one of 18 reference files, from core-directory-structure.md to best-practices-ssr.md. Adding --recurse-submodules matters only if you plan to regenerate rather than install, since sources/ holds the upstream documentation checkouts.

What It Provides¶

Skill	Installs	Use For
tailwindcss	3,884	Utility-first styling: responsive layouts, theme variables and design systems.
hairy-utils	1,258	Working with the @hairy/utils core utility package.
motion	1,148	Animations, gestures, layout transitions and scroll effects with Motion for JS, React and Vue.
openapi-specification-v3.2	906	Authoring and validating OpenAPI 3.2 documents, including JSON Schema 2020-12.
next	858	Next.js App Router, Server Components, caching and Server Actions.
react-use	848	Picking hooks for sensors, UI, animation, side effects and state.
vitest	786	Writing tests, mocking, coverage configuration and test filtering.
valtio	784	Proxy-based reactive state for React and vanilla JavaScript.
openapi-specification-v2	779	Writing or validating Swagger 2.0 specs and generating clients and docs.
valtio-define	753	Working with the valtio-define helper package.
hairy-react-lib	752	Using the @hairy/react-lib hooks and components.
unjs	732	h3, nitro, ofetch, unstorage and other universal JavaScript builds.
overlastic	678	Working with the Overlastic library.
hairy	657	Hairyf's opinionated defaults for TypeScript, ESLint, pnpm and Vitest.
undocs	630	Documentation sites built with Nuxt Content and Nuxt UI.
writing-humanizer-zh	627	Removing AI-writing tells from Simplified Chinese copy.
tsdown	619	Bundling TypeScript libraries with Rolldown and emitting type declarations.
arch-tsdown-monorepo	611	Scaffolding a pnpm monorepo with tsdown per package and trusted publishing.
tauri	415	Tauri apps: Rust backend, WebView frontend, IPC, security and plugins.
react-native-reusables	377	shadcn-style components for React Native with NativeWind or Uniwind.

The remaining 60 indexed listings range from 2 to 374 installs.

The generated skills share one shape: frontmatter with name, a trigger-heavy description, and metadata recording author, version and source repository, followed by a table of topics that points into references/. That is why the collection can be large without being expensive, because an agent loads one SKILL.md of a few hundred lines and follows only the reference links its task needs. The hand-maintained skills listed in the README (hairy, antfu, github-workflow, create-skill-from-repo, dsh-plugin-creator, the two OpenAPI specs and the media skills) are prose rather than generated, so they read differently: fewer reference tables, more direct instruction.

Why This Matters for Hermes Agents¶

The context budget is the whole argument for this format. Nearly 2,000 Markdown files would be impossible to carry into a session; as skills they cost one description each until a request matches, and the README makes that point itself when it contrasts skills with an AGENTS.md that loads everything upfront.

Generated skills carry provenance. The nuxt skill states that it is based on Nuxt 3.x and was generated on 2026-01-28, with metadata naming the upstream repository. An agent that knows which framework generation it is reading is less likely to suggest an API that was removed two majors ago, and a stale skill is visible as stale rather than merely wrong.

The arch-* skills are the part with no equivalent in a documentation set. They are opinions about how to scaffold a Nuxt app, a tsdown library, a CLI, a monorepo, a VSCode extension or a browser extension, which is exactly the kind of decision an agent makes badly by default and a reviewer then pays for.

The risk is overlap. Eighty skills mean many near-duplicate triggers across the Vue and React ecosystems, and the router, not the model, decides which one loads. Installing all of them into one profile is how an agent ends up reading the React skill for a Vue request.

Usage¶

You say	What happens
"Set up linting and formatting for this new TypeScript library"	`hairy` or `antfu` applies the opinionated defaults: pnpm, @antfu/eslint-config, Vitest, strict TypeScript.
"Scaffold a lightweight Vue SPA"	`arch-nuxt-lite` lays out the Vitesse Lite stack: Vite, Vue 3, file-based routing, UnoCSS and VueUse.
"Write tests for this composable"	`vitest` covers config, mocking, coverage and test filtering in a Vite project.
"Publish this package as ESM with type declarations"	`tsdown` bundles with Rolldown and emits declarations, with the `arch-tsdown` starter for layout.
"Build a Nuxt module with runtime code and types"	`arch-nuxt-module-builder` scaffolds with @nuxt/module-builder.
"Move this React Native app from NativeWind to Uniwind"	`migrate-nativewind-to-uniwind` handles package removal, config migration, Tailwind 4 and theme conversion.
"Make this Chinese post sound less like a model wrote it"	`writing-humanizer-zh` rewrites Simplified Chinese copy against a catalogue of AI-writing patterns.

Verification¶

The raw paths below returned 200 on branch main on 2026-11-03:

curl -sL -o /dev/null -w "%{http_code}\n" \
  https://raw.githubusercontent.com/hairyf/skills/main/skills/tailwindcss/SKILL.md

curl -sL -o /dev/null -w "%{http_code}\n" \
  https://raw.githubusercontent.com/hairyf/skills/main/skills/nuxt/SKILL.md

Then count what a clone actually delivers, and inspect one generated skill:

git clone --depth 1 https://github.com/hairyf/skills.git /tmp/hairyf-skills
ls -d /tmp/hairyf-skills/skills/*/ | wc -l
find /tmp/hairyf-skills/skills -name SKILL.md | wc -l
ls /tmp/hairyf-skills/skills/nuxt/references | wc -l
head -8 /tmp/hairyf-skills/skills/nuxt/SKILL.md

The counts should print 80, 80 and 18. head should show name: nuxt, the description with its trigger phrases, and metadata naming Anthony Fu as author with a version stamp. If a skill directory holds only SKILL.md and no references/, the copy came from a partial checkout and the skill will answer from headings alone.

Security¶

skills.sh verdicts for sampled skills (verified 2026-11-03) - re-check the security pages on skills.sh before production use:

Skill	Gen Agent Trust Hub	Socket	Snyk
tailwindcss	Pass	Pass	Pass
hairy-utils	Pass	Pass	Pass
motion	Pass	Pass	Pass

The three sampled listings are documentation only, and most of the tree is Markdown, but the repository as a whole is not inert. skills/vision/scripts/vision.js is a Node CLI that sends images to an OpenAI-compatible vision model, skills/sonic/scripts/install/woosh.sh and mmaudio.sh install local audio tooling, skills/minecraft-texture/scripts/mc_texture.py post-processes images into block textures, and skills/create-skill/scripts can initialise and package new skills. None of that runs until a skill is invoked, but installing all 80 copies all of it into the profile, and the green tier reflects three sampled listings rather than an audit of every script.

The content is also third-hand in places: many skills are generated from upstream documentation repositories cloned as submodules, and the README credits further sources such as vercel-labs/agent-skills, blader/humanizer and uni-stack/uniwind. Check the upstream terms before redistributing a generated skill, and treat sonic and vision as network clients that need keys rather than as pure knowledge.

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

Limitations¶

Snapshot data, verified 2026-11-03: 25,708 combined installs across 80 indexed listings; 26 GitHub stars; MIT via LICENSE.md; last pushed 2026-10-14. Counts drift over time.
Install counts and repository attention disagree sharply: 25.7K installs against 26 stars. Distribution comes through skills.sh, so treat stars as a weak signal here.
The README calls the collection a proof of concept and says the author has not fully tested how well the skills perform in practice. Read a skill before you rely on it.
Generated skills are only as current as the commit they were generated from. The nuxt skill is based on Nuxt 3.x as of 2026-01-28, and nothing in the install path refreshes it.
Several skills are Chinese-first (uniapp, uniapp-x, writing-styles-juejin, writing-humanizer-zh) and produce or expect Chinese-language output.
The 60 listings below 377 installs include most of the specialised material (Minecraft, Electron, Flutter, the smaller arch starters), so the table above is the popular half rather than the interesting half.

Skills Catalog | Skills Marketplace

Pro Tips¶

Install the framework skill plus its arch-* starter, not the whole collection. Two or three skills cover a normal Vue or Nuxt project.
Check references/ before trusting a generated skill. A directory with four reference files is a summary; one with eighteen is close to the upstream docs.
Read the metadata.source and generation date when a framework ships a major version, and regenerate rather than patch by hand.
The fork is a working template for the same trick in another ecosystem: meta.ts lists source repositories, and the README documents nr start init, sync and cleanup as the maintenance loop.
Prefer the hand-maintained skills for conventions (hairy, antfu, github-workflow) and the generated ones for API surface, because the two are maintained in different ways.

← Skills Catalog | Skills Marketplace → Powered by CorpusIQ