Show HN: SkillWorks – every Claude Code skill, scored on whether it loads

SkillWorks, a nightly index of Claude Code skills, has scored 503,570 listings and found 48,190 that fail to load, according to data rebuilt on 2026-08-22. The project, which reads 23,438 repositories, scores each artifact 0–100 based on four weighted components: works (40%), maintained (25%), adopted (20%), and docs (15%).

Every Claude Code skill, scored Read from the source and scored 0–100 on four weighted components, every night. 503,570 listings · 392,227 skills · 80,636 subagents · 23,442 plugins · 7265 marketplaces · 6,267 with a tracked install count · 48190 that do not load · rebuilt 2026-08-22 | Artifact | Score | Type | works 40 | maint 25 | adopt 20 | docs 15 | Reach | Last commit | |---|---|---|---|---|---|---|---|---| | find-skills /skills/vercel-labs/skills/find-skills vercel-labs/skills ui-ux-pro-max /skills/nextlevelbuilder/ui-ux-pro-max-skill/ui-ux-pro-max nextlevelbuilder/ui-ux-pro-max-skill pptx /skills/anthropics/skills/pptx anthropics/skills docx /skills/anthropics/skills/docx anthropics/skills mcp-builder /skills/anthropics/skills/mcp-builder anthropics/skills turborepo /skills/vercel/turborepo/turborepo vercel/turborepo vercel-optimize /skills/vercel-labs/agent-skills/vercel-optimize vercel-labs/agent-skills sandbox-bench /skills/vercel/next.js/sandbox-bench vercel/next.js microsoft-foundry /skills/microsoft/azure-skills/microsoft-foundry microsoft/azure-skills azure-quotas /skills/microsoft/azure-skills/azure-quotas microsoft/azure-skills Browse all 504k /search Every row links to the checks it passed, the checks it failed, and the facts each component was computed from. The index, in numbers 2026-08-22Listings scored 503,570 2026-08-22Skills, subagents, plugins and marketplace repositories, every one of them read from the file rather than from a list somebody maintains.Repositories read 23,438 Every repository opened and read so far, across all runs to date. Found through GitHub topic and code search, the skills.sh registry, the large awesome-lists, and anything submitted here.Will not load 48,190 2026-08-22Each one fails a check Claude Code needs to register it at all. Published in full at /broken /broken rather than quietly dropped. With a tracked install count 6,267 2026-08-22Real installs from the skills.sh registry. Everything else falls back to repository stars, and says which of the two it is showing. Four things Claude Code can load Each is a different file shape with a different set of ways to be broken, so each is indexed and scored separately. 392,227 Skills Folders of instructions Claude Code loads on demand, defined by a SKILL.md file. 80,636 Subagents Named agents with their own prompt, model and tool allow-list, defined by a markdown file under agents/. 23,442 Plugins Packaged bundles of skills, agents, hooks, MCP servers and LSP servers, installed from a marketplace repo. 7,265 Marketplaces Repos that publish a .claude-plugin/marketplace.json at their root, which Claude Code adds as a plugin source with /plugin marketplace add. What a README list cannot tell you The alternatives are an awesome-list nobody re-checks and a scraper that counts stars. Both answer “does this exist”. Neither answers “will this load”. | The question | An awesome-list README | SkillWorks | |---|---|---| | Does the frontmatter parse? | Not checked | Checked on every artifact, every night | | Are the files it references in the repository? | Not checked | Resolved, and listed by name where they are missing | | How many are broken right now? | Unknown | 48190 of 503,570, published at | How the score is built The full method /methodology Four weighted components, all four read straight from the repository. No execution, no model judging anything, no hand curation, and no way to buy a point. 40% of the score Works Does it parse, does it declare a name and a description, does it reference files that are actually there. A failure here is what sets the will-not-load verdict. 25% of the score Maintained Days since the last commit on the repository that publishes it, on a steep curve, and whether that repository has been archived. 20% of the score Adopted Install counts from the skills.sh registry where they exist; repository stars and 30-day star velocity where they do not. 15% of the score Documented Length of the instructions, worked examples, bundled scripts and references — how much a model actually has to go on. 48190 of them will not load All 48190 /broken Popular, starred, and structurally broken. This is the half of the index the other directories do not publish, and it is most of the reason this one exists — these are the six with the largest audiences. | Artifact | What stopped it | Reach | |---|---|---| | vercel-react-native-skills /skills/vercel-labs/agent-skills/vercel-react-native-skills vercel-labs/agent-skills convex-create-component /skills/get-convex/agent-skills/convex-create-component get-convex/agent-skills momentic-result-classification /skills/momentic-ai/skills/momentic-result-classification momentic-ai/skills clerk-setup /skills/clerk/skills/clerk-setup clerk/skills clerk-nextjs-patterns /skills/clerk/skills/clerk-nextjs-patterns clerk/skills By what it is for All categories /categories Assigned from the artifact’s own name, description and repository topics by deterministic keyword rules, applied fresh every night. Counts are of listings that pass every structural check. Nobody else re-reads the files Every list of Claude Code skills was assembled once and has been decaying ever since. A repository is archived, a frontmatter block is edited into something that no longer parses, a referenced file is renamed — and the list recommending it stays exactly where it was. Nothing on it says whether any of it still works, because nothing on it ever checked. This index reads the file. Every SKILL.md, every agent markdown file, every plugin and marketplace manifest is fetched from the repository that publishes it and put through the same structural checks Claude Code’s own loader applies, once a night. 48190 of the 503,570 listings here fail one of them, and they stay listed, with the check they failed named on the row — a directory that hides its own bad results is a list again. The score is four measured components and nothing else. No model is asked whether a skill is good, nothing is executed in a sandbox, nobody curates, and sponsorship sits outside the ranking and cannot move a number. All of that is written down at /methodology /methodology before it says anything else.