One Repo, 92,647 Stars, Four Skills — What Megarepos Actually Ship
Star count is not a score — we've said that before. But the megarepos are interesting for a different reason: when a repository with 25k–92k stars ships agent skills, what do those skills actually contain? We went through the top four by star count and found something unexpected — they're not wrappers.
Our directory's top scores are dominated by household names in the skill world: the Office trio, the API references, the HyperFrames family. But sort by source-repository stars and a different list appears — repositories with 25,000 to 92,000 GitHub stars, each shipping a small number of agent skills alongside their main project.
The star counts are inflated by the parent projects, not the skills. That's exactly why they're worth examining: with the hype variable maxed out, what remains is the actual skill engineering.
The four megarepos, by what they ship
Repo: 92,647 stars — four skills, all 9.4+
constraint-driven-development— 9.5: establishes a project's quality bar as a written contract, then holds every change to it. A methodology skill, not a tool.interview-me— 9.4: extracts what the user actually wants instead of what they say. We've written about this pattern before — it's one of the best interview skills in the directory.archify— 9.4: polished, validated architecture and workflow diagrams.- plus a UI-extraction skill for reverse-engineering interfaces.
Repo: 56,452 stars — the HyperFrames family, ten skills
Covered in its own post: a complete programmatic-video pipeline, 8.7–9.7 across nine skills, one outlier at 5.7. The most designed skill system in the directory.
Repo: 25,617 stars — the baoyu family, twenty-one skills
A single author's personal toolkit covering translation, diagrams, image generation, WeChat publishing, YouTube transcripts, slide decks, infographics — 8.9 to 9.5, no filler. The largest coherent family we host.
Repo: 30,919 stars — one skill
vercel-optimize — 9.4, data category. A performance-optimization analyzer. The narrowest megarepo skill, and proof that one well-scoped skill can carry a big repo's agent story by itself.
What the megarepos do differently
Three things, none of which require 92k stars to imitate:
1. They ship methodologies, not wrappers. The 92k repo's best skill (constraint-driven-development, 9.5) contains no API calls, no tools, no integrations — it's a way of working, written precisely enough that an agent can hold a project to it. When a megarepo ships a skill, it's usually the repo's thinking, condensed.
2. Small counts, complete coverage. Four skills, ten skills, one skill. No megarepo ships forty — they pick the jobs where an agent genuinely helps their users and stop. Contrast with the average multi-skill repo we evaluate, which often ships fifteen near-identical capability wrappers.
3. The skills stand alone. Read each SKILL.md without knowing the star count and you'd still score it 9+. The skill doesn't lean on the parent project's reputation; it defines its own trigger, scope, and verification. That's the trigger field problem solved by discipline, not distribution.
What stars actually correlate with
Nothing we can measure — which matches our earlier finding. Across the directory, star count and our score have essentially no relationship above ~2,000 stars. What the megarepos demonstrate instead is a selection effect: teams disciplined enough to build a 90k-star project tend to write disciplined skills. The stars don't transfer. The discipline does.
The practical takeaway for skill authors: you can't borrow a megarepo's distribution, but the recipe is fully copyable — pick one job where you genuinely know the methodology, write it as a contract your agent can hold, scope it tightly, and stop.
Where to browse
All four repos' skills are in the directory with full scorecards — start from the top of the leaderboard and follow the star badges, or filter by developer tools and productivity.