antfu/skills
The Vue ecosystem's most prolific maintainer ships a one-stop skill pack, plus the machinery to generate your own from upstream docs.
The cleanest multi-skill collection GearScope has parsed: strict YAML passes on all 11 SKILL.md files and all 139 references, installs land byte-identical, and the generator machinery is self-consistent. The catch is freshness: 5 of the 9 doc-generated skills still sit at January baselines, so pinia and pnpm guidance runs a major version behind npm reality. If you work in the Vue/Vite/Nuxt stack this is a KEEP IT today, with a calendar note to re-check the generated half.
$pnpx skills add antfu/skills --skill='*'
$npx -y skills add antfu/skills --skill='*' --agent hermes-agent --copy -y
$npx -y skills add antfu/skills --skill vue --agent claude-code --copy -y
install if
- Vue/Vite/Nuxt-stack developers using coding agents daily. Eleven skills covering the exact toolchain (plus the author's setup conventions) with demand-proven install paths and byte-exact delivery is the strongest single collection for this stack.
- Teams standardizing on antfu-style conventions (ESLint config, pnpm catalogs, pure ESM, UnoCSS tokens). The antfu and antfu-design skills encode those preferences in reusable, referenced form.
- Skill authors who want to maintain doc-derived skills for their own projects. The submodule-plus-GENERATION.md pattern makes staleness auditable, and the whole repo is a working template.
skip if
- Anyone outside the JavaScript/Vue toolchain. Ten of eleven skills are Vue-ecosystem-specific; there is nothing here for Python, Go, or mobile stacks.
- Users who need version-current accuracy right now. pinia and pnpm guidance runs a major version behind npm, and nitro documents the v3 beta; check GENERATION.md dates before relying on API details.
- Shoppers wanting evals or activation guarantees. The README itself frames the collection as an untested proof of concept; there is no harness measuring whether agents actually pull the right skill at the right time.
What It Does
antfu/skills is Anthony Fu's curated Agent Skills collection for the Vue/Vite/Nuxt toolchain: 11 skills covering his personal conventions (antfu, antfu-design) plus nine skills generated from the official documentation of Vue, Nuxt, Nitro, Pinia, Vite, VitePress, Vitest, UnoCSS, and pnpm through git-submodule snapshots of each docs repo. Each generated skill is a small SKILL.md index (45 to 84 lines) routing to per-topic reference files (139 files, 26,384 lines total). The repo doubles as a framework: a 408-line manager CLI (init, sync, check, cleanup), a meta.ts manifest of doc sources, and a 207-line AGENTS.md spec that tells an agent how to regenerate skills when upstream docs move. The morning of this review, PR #39 removed the entire vendored-skills concept (eight third-party synced skills deleted), shrinking the collection from 19 to 11 skills. The author states plainly in the README that this is a proof of concept that he has "not fully tested... in practice."
The Good
Install fidelity is exact, on two agents, with lockfile receipts. Running the README's install command in scripted form landed all 11 skills for hermes-agent; every landed SKILL.md was byte-identical to the repo clone, all 139 reference files arrived, and skills-lock.json was written. A single-skill request (vue, claude-code) landed exactly one skill, also byte-identical. Requesting a skill name that today's PR removed (slidev) failed loudly with exit 1 and a list of the 11 available names rather than installing something stale.
Structural hygiene is the best GearScope has measured in a multi-skill pack. A strict PyYAML parse of all 11 SKILL.md files and all 139 reference files produced zero errors, every frontmatter name matches its directory, all 142 internal reference links resolve (four initial hits were link-syntax examples inside code fences, not real links), the repo's own eslint config passes, and node scripts/cli.ts cleanup -y reports the tree is clean, confirming the manifest matches reality after the same-day cull. For comparison, Unity's official pack shipped 4 of 31 skills with unparseable frontmatter.
Staleness is auditable by design, which is rare and worth copying. Every generated skill ships a GENERATION.md recording the source repo SHA and date it was generated from, and the submodule pointers sit in the tree, so anyone can compute drift without trusting the author. The audit found the June 2026 regeneration pass (nitro, nuxt, pnpm, vitest) in sync and the January cohort (pinia, unocss, vite, vitepress, vue) behind their own moved pointers. The antfu-design skill also credits its influences in a source block, naming Leonxlnx/taste-skill and other repos it distilled from.
Demand is real and measurable. skills.sh exact-name probes put eight live entries under the antfu/skills repo at roughly 125,000 combined installs (vitest 37,471, pnpm 21,109, pinia 18,598, antfu 16,187, unocss 15,587, vitepress 11,405, nitro 2,100, antfu-design 2,055), for a repo created eight months ago with 5,923 stars, 334 forks, and 61 commits from 11 contributor names including eight outside the author's accounts.
The Bad
The keep-them-in-sync thesis is only half realized, and two skills are a major version behind. Five of the nine doc-generated skills (pinia, unocss, vite, vitepress, vue) still carry January 2026 generation dates while their source submodule pointers have moved on. Concretely: the pinia skill says "based on Pinia v3.0.4" while npm serves 4.0.3, and the pnpm skill covers 10.x with v11 notes while npm serves 12.6.0. The nitro skill documents the v3 beta while the nitropack dist-tag latest is 2.13.4, so stable-line users get beta-flavored guidance. The vue, vite, vitepress, and unocss baselines all still match current majors, so the drift is uneven rather than uniform.
The README trails the repo it documents. The skills tables list 10 of the 11 shipped skills: nitro is installed by every --skill='*' invocation but never mentioned. instructions/tsdown.md still sits in the tree with generation preferences for a tsdown skill that PR #39 deleted hours before this review. There are no git tags, no releases, and no changelog at 5.9K stars, so consumers have no version to pin and no signal for culls like today's beyond watching main.
Registry surfaces lag and collide. Five of the eight skills deleted today (slidev, tsdown, vueuse-functions, vue-router-best-practices, vue-testing-best-practices) were still listed on skills.sh search at review time with 86,053 combined installs; the installer is safe because it resolves against the live repo, but the catalog pages mislead. The three most generic names (vue, vite, nuxt) do not surface an antfu-owned entry in top search results because other publishers own same-name skills there, which matters for discovery by name alone.
The fork-and-generate path is heavy and unproven. A plain pnpm install triggers a prepare hook that recursively clones all nine doc-source submodules; the seven whose sizes the API would still serve total about 215 MB before counting the nuxt and vite monorepos. Regeneration then depends on an interactive agent session following AGENTS.md, with no eval harness, no activation tests, and the author's own proof-of-concept caveat.
Smoke Test Results
Host-based runs on macOS 27.0 (aarch64), Node 24.13.1, corepack pnpm 10.32.1 (the sbx microVM daemon is down pending Docker auth, so the documented host fallback was used). Three variants from clean clones; logs are in the sandbox folder.
Run A: Fresh sandbox, no deps preinstalled
$ git clone --depth 1 https://github.com/antfu/skills.git && du -sh .
β
1.6M total
$ find skills -name SKILL.md | wc -l
β
11
$ find skills -path "*/references/*.md" | wc -l
β
139
$ for f in skills/*/SKILL.md; do head -1 "$f" | grep -q '^---$' && sed -n '2,10p' "$f" | grep -q '^name:'; done; echo rc=$?
β
rc=0, all 11 carry name+description frontmatter
$ node scripts/cli.ts cleanup -y
β
exit 1, ERR_MODULE_NOT_FOUND on @clack/prompts: the CLI fails closed with a clear error until devDependencies are installed
Pass rate: 5 of 5. The last check passes as documented fails-closed behavior, not a defect.
Run B: Sandbox with deps preinstalled
$ python (PyYAML venv): strict parse of 11 SKILL.md + 139 references
β
zero YAML errors, every name matches its directory
$ python: cross-reference resolution, code fences excluded
β
142 internal links checked, 0 broken
$ python: README claims vs shipped skills
β skills shipped but absent from README tables: ['nitro']
$ python: GENERATION.md SHA vs submodule pointer audit
β
audit computed: 4 in sync (nitro/nuxt/pnpm/vitest, June 22), 5 drifted (pinia/unocss/vite/vitepress/vue, January)
$ corepack pnpm install --ignore-scripts
β
7 packages installed, exit 0
$ node scripts/cli.ts cleanup -y
β
exit 0, "Everything is clean, no unused submodules or skills found"
$ node scripts/cli.ts check
β
exit 0 (vacuous without submodule clones; documents the workflow entry)
$ corepack pnpm lint
β
eslint (repo's own @antfu/eslint-config) passes clean
Pass rate: 7 of 8. The one failure is the nitro README omission, a real documentation defect, not a checker artifact (the initial run's four broken-link hits and the sync-audit error were bugs in our first checker, fixed and re-run; both are gone in the final log).
Run C: Functional verification
$ npx -y skills@latest add antfu/skills --skill='*' --agent hermes-agent --copy -y (isolated HOME)
β
11 skills landed; all 11 SKILL.md byte-identical to the clone; 139 references landed; skills-lock.json written
$ cmp .hermes/skills/$s/SKILL.md repo/skills/$s/SKILL.md (per skill)
β
0 byte differences across all 11
$ npx -y skills@latest add antfu/skills --skill vue --agent claude-code --copy -y
β
exactly 1 skill landed, byte-identical
$ curl skills.sh/api/search?q={name} (11 per-name probes)
β
8 of 11 names resolve to antfu-owned entries, ~125K combined installs; vue/vite/nuxt buried under other owners' same-name skills
$ curl skills.sh/api/search?q={removed-name} (8 probes against PR #39's deletions)
β
5 ghost entries still served (86,053 installs); live install of slidev fails loudly: "No matching skills found for: slidev", exit 1
$ npm view {pkg} version vs skill baselines
β
baselines match current majors for vue/vite/vitepress/unocss/nuxt/vitest; pinia (3.0.4 vs 4.0.3) and pnpm (10.x+11 notes vs 12.6.0) are a major behind
Functional pass rate: 6 of 6. Passes mean the probes executed and returned usable evidence; the drift findings they surfaced are reported in The Bad.
What the runs tell you
Installation and structure are production-grade: everything lands byte-identical, everything parses, the manifest machinery agrees with the tree. The risk concentrates entirely in content freshness of the generated half and in documentation trailing the repo by one skill. Buy it for the stack coverage and the conventions; re-audit the generated skills' GENERATION.md dates before trusting version-sensitive details.
Setup Walkthrough
- Install everything:
pnpx skills add antfu/skills --skill='*'(README-verbatim). In scripts or CI, usenpx -y skills add antfu/skills --skill='*' --agent hermes-agent --copy -y; the picker-free form needs the agent and copy flags to avoid stalling without a TTY. - Pick individual skills instead:
npx -y skills add antfu/skills --skill vue --agent claude-code --copy -ylands exactly one skill. Useful when you want pnpm and vitest but not the design opinions. - To fork it as a generator template: clone, edit meta.ts with your doc sources,
pnpm install(expect the prepare hook to start pulling nine doc submodules, several hundred MB all told), thenpnpm start init,pnpm start sync, and ask your agent to generate skills per AGENTS.md, one project at a time. - Post-install gotcha: there are no releases, so installs track main. Today that meant the collection shrank from 19 to 11 skills mid-day (PR #39). Pin by commit in skills-lock.json if you need reproducibility.
Alternatives
- onmax/nuxt-skills: a Nuxt-only registry rival whose
nuxtentry outranks antfu's in skills.sh search; prefer it for Nuxt depth, though it covers one tool instead of nine. - mattpocock/skills: an opinionated TypeScript workflow framework (grill, spec, tickets, implement, review) that complements rather than replaces this pack; choose it for process discipline over library knowledge.
- Leonxlnx/taste-skill: the 80K-star design-taste prompt-contract pack that antfu-design partially credits as a source; broader design coverage with no toolchain coupling.
Reviews stay honest because nobody pays us to publish them. If this one saved you time, throw a coin.
Tip the reviewer- reviewed by
- GearScope
- tested
- 2026-09-25 · macOS (Apple Silicon)
- last verified
- 2026-09-25
- depth
- HANDS-ON
- sponsorship
- none, ever
Don't install your next skill blind.
Every week: the shortlist of skills worth installing β and the ones to skip β from 100+ hands-on tests. No spam, no affiliate links.