Meta-Skills Roundup

The list first: nine official items, with install commands

How to install (commands quoted from the official pages)

# ① The official skills repo: register the marketplace, then install by bundle
/plugin marketplace add anthropics/skills
/plugin install example-skills@anthropic-agent-skills
/plugin install document-skills@anthropic-agent-skills

# skill-creator also has an official plugin source (this is the line in the Claude Code docs)
/plugin install skill-creator@claude-plugins-official

# ⑧ Audit: context cost and invocation frequency per skill (needs v2.1.252+)
/skill-doctor

# ⑨ Validate: check YAML and manifest syntax
claude plugin validate .claude/skills

Item by item

• ① skill-creator (official): the real payload is the second half. You write 2–3 realistic test prompts, then it spawns parallel subagent runs with and without the skill, grades the outputs against your assertions using agents/grader.md, runs scripts.aggregate_benchmark into benchmark.json, and opens eval-viewer/generate_review.py for a human pass. The Claude Code docs describe the same chain as: cases stored in evals/evals.json inside the skill directory → one isolated subagent per case → results into grading.json → pass rate, time and tokens aggregated into benchmark.json. It also does blind A/B comparison and description tuning.
• ② mcp-builder (official): splits "write an MCP server" into research and planning → implementation → review and testing → evaluation, with separate implementation guides for Python (FastMCP) and Node/TypeScript. The hard requirement is ten evaluation questions, each independent, read-only, complex, realistic and verifiable — and you must solve each one yourself first, with answers that support string comparison.
• ③ web-artifacts-builder (official; note the real name is not artifacts-builder): a toolchain that bundles React 18 + TypeScript + Tailwind + shadcn/ui into a single self-contained HTML file. Every deterministic step is a script — scripts/init-artifact.sh scaffolds the project, scripts/bundle-artifact.sh uses Parcel to produce bundle.html — and the scaffold ships with 40+ shadcn/ui components pre-installed.
• ④ claude-api (official): detects the project's language first, then reads only that language's docs; shared material is split by domain into shared/prompt-caching.md, shared/cost-optimization.md, shared/model-migration.md, shared/token-counting.md and more. It isn't a skill-building tool; it's here because it is the most complete official specimen of "one SKILL.md indexing many domain files" (Skill 23 takes it apart).
• ⑤ template/SKILL.md: the repo's template/ folder holds exactly one file, and it is almost provocatively short — frontmatter with name: template-skill plus one line telling you to replace the description with what the skill does and when Claude should use it, then a single body line, # Insert instructions below. That is Anthropic's idea of a minimum viable skeleton.
• ⑥ agent-skills-spec: spec/agent-skills-spec.md in the repo is now one sentence — the spec has moved to agentskills.io/specification. That fact is itself the intelligence: the source of truth for skill fields has left Anthropic's repo for an independent site (Claude Code's skills page points its evaluation chapter at the same domain). Copy a 2025 snapshot of the repo and you are copying a stale spec.
• ⑦ Skill authoring best practices (platform.claude.com): the origin of every criterion in section 03, including the five-step "build evaluations first" method and the closing checklist.
• ⑧ /skill-doctor and ⑨ claude plugin validate: the first lists context cost and usage frequency per skill and flags the ones never invoked plus plugins unused recently; the second checks marketplace.json for schema errors, duplicate plugin names, source path traversal and version mismatches.

Don't rank these nine on one list: the denominators differ

Background

anthropics/skills shows roughly 175k stars and 20.7k forks per the GitHub repo page, 2026-09 (this site's Skill 01 recorded 172k in 2026-08 — about 3k in a month). But that number measures how famous the demo repo is, not which skill inside it anyone uses. All 19 skills share one star count, which makes skill-creator and slack-gif-creator exactly equivalent on that metric. Obviously absurd.

The more practical reason to keep the layers apart is substitutability. ⑤ the template and ⑦ the best-practices page are plain text: copy them into your own repo and you own them forever, with no version risk. ①—④ are executable, so upgrades, dependencies and script rot follow upstream. ⑧—⑨ are Claude Code capabilities and vanish the moment you switch harness. Three different shelf lives; don't shop for them with one rule.

In practice

One more trap: names get copied wrong. ③ is really web-artifacts-builder, and plenty of aggregator articles print artifacts-builder — install by that name and it fails; point an agent at it and it finds nothing. ⑥ has moved in the same way: the file is still in the repo, but its content is now a single line saying the spec lives at agentskills.io. Standing rule: any name or path you write into a config or hand to an agent comes from the repo directory or the official doc page, never from an aggregator.

Three-minute inspection: judge candidates by how the official four are built

Background

First the hard limits, quoted from the authoring page: name is at most 64 characters, lowercase letters, digits and hyphens only, and may not contain reserved words like anthropic or claude; description is at most 1,024 characters, non-empty, and must be written in the third person; the body should stay under 500 lines; references must be one level deep. The Claude Code docs add a number that is easy to miss: in skill listings, description plus when_to_use is truncated at 1,536 characters — the overflow isn't read late, it's cut.

In practice

1. Does the description carpet the trigger space? Read skill-creator's own: "create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description." Five phrasings in one sentence, because the user could say any of them. A description reading "Helps with skills" either never fires or fires everywhere.
2. Is there a machine-checkable acceptance bar? mcp-builder requires those ten questions to have answers that "support string comparison verification." That sentence is the watershed: an author who writes it knows what they're aiming at; one who can't, ships to a standard of "looks fine."
3. Are deterministic steps frozen into scripts? web-artifacts-builder puts scaffolding and bundling into init-artifact.sh and bundle-artifact.sh instead of regenerating them each run. Anthropic's stated reasons are blunt: pre-made scripts are more reliable, cheaper (only the output is billed; the script body never enters context) and consistent.
4. Are references split by domain, one level deep? claude-api splits into a dozen-plus files by language and topic, leaving SKILL.md as a table of contents. The docs warn explicitly that with nested references Claude may preview with something like head -100 and act on partial information — so every reference links straight from SKILL.md, and reference files over 100 lines carry their own table of contents.
5. Does the freedom match the task's fragility? The official analogy is a narrow bridge with cliffs on both sides versus an open field. Database migrations want low freedom (an exact command, plus "do not modify the command or add additional flags"); code review wants high freedom (direction only).
6. Can it name the three concrete failures it fixes? The official checklist says "at least three evaluations created" and "tested with Haiku, Sonnet, and Opus." A meta-skill that can't answer this leaves you no way to tell whether it helps.

And don't invert the order. The authoring page makes "create evaluations BEFORE writing extensive documentation" a bolded requirement, in five steps: run real tasks with no skill and document the specific failures → build three scenarios around those gaps → measure the baseline → write just enough instruction to close them → iterate against the baseline. The stated reason: this way you solve real problems instead of documenting imagined ones.

The honest cost: the docs themselves say you don't need a skill-writing skill

Background

There is a Tip on the authoring best-practices page that is deflating in its directness: Claude understands the skill format natively; you don't need special system prompts or a "writing skills" skill — just ask Claude to create a skill and it generates properly structured SKILL.md content with appropriate frontmatter.

So where is skill-creator's value? In the evaluation machinery of its second half, not the drafting. Parallel with/without runs, assertion grading, an aggregated benchmark.json — building that by hand costs you a day. Install it expecting "generate my SKILL.md" and you've paid a context tax for something that was already free.

In practice

• The evaluation story is forked; one file does not run everywhere. The platform docs give an eval JSON structure (skills / query / files / expected_behavior) and then state plainly that there is not currently a built-in way to run these evaluations — you build your own runner. The Claude Code side's skill-creator has its own evals.json → grading.json → benchmark.json pipeline. The two are not interchangeable.
• Passing validate ≠ a healthy skill. The docs are explicit: run from a marketplace directory, Claude Code does not open the plugins' skill, agent, command or hook files; to find errors in those you must validate the individual plugin. Treating claude plugin validate as a quality gate is a misread — it is a syntax check, not a content check.
• Meta-skills are permanent residents installed for occasional work. You write five skills a year and let skill-creator sit in your description candidate pool for all twelve months. So audit after you ship: /skill-doctor flags skills that were never invoked. Delete those — keeping them only makes the rest harder to select.
• Ecosystem drift is a real cost. The spec moved to agentskills.io and Claude Code's evaluation chapter points there too, which means your skill library now tracks a version outside the repo. Teams working across harnesses should copy the current spec's key fields into their own repo rather than looking them up fresh each time.