September 7, 2026SkillsOpen SourceCoding

A Quarter Million Stars for One Engineer's .agents Folder

The number two repo on GitHub trending today, gaining 2,206 stars in a single day on top of 254,000 total, is a folder of markdown files. mattpocock/skills (https://github.com/mattpocock/skills) is Matt Pocock — the TypeScript educator with a 60,000-subscriber newsletter — publishing his personal .agents directory under the tagline "Skills for Real Engineers." MIT licensed, works with Claude Code and Codex.

The contents read like scar tissue from a year of agent-driven development. /tdd enforces a test-first loop. /code-review runs a two-axis review discipline. /grill-with-docs interviews you to build a domain model before writing any code. /handoff compresses a conversation so the next agent can pick up where this one died. The README names the four failures these exist to prevent: misalignment with requirements, verbose output, broken code, architectural decay. The philosophy is explicitly small, composable, hackable — skills you can read in one sitting, against frameworks you can't.

Look at the rest of the board and it stops being one repo's story. humanlayer/skills (+451 today) ships skills like design-control-loop and build-iterated-agentic-loop, installed via npx skills add. openai/skills, the official Codex catalog, sits at 25,600 stars. coreyhaines31/marketingskills — CRO, SEO, copywriting for Claude Code — is at 47,000 and still climbing. The trending page has quietly become a skills shelf, and it's expanding past engineering into marketing, with every other profession queued up behind.

The supply chain now has two ends. BAAI distills a thousand GitHub repos into five thousand skills by machine (https://clauday.com/article/6d566938-69a4-4a25-b026-01b6def425be); Pocock hand-writes twenty from personal pain. A quarter million stars for the artisanal end is a real signal about which one developers currently trust — expertise as installable text is the distribution moat of this cycle, and the people who've actually felt the failures are winning it so far.
← Previous
OpenAI's Chief Scientist: No Lab Has Solved Alignment
Next →
magnitude's Answer to Agent-Loop Costs: $0, Use Your Own GPU
← Back to all articles

Comments

Loading...
>_