I kept seeing the same agent-skills repos on GitHub Trending.
So I cloned six of them.
One had 33,000 stars and four files. Another was 98% machine translation. A third wrote down engineering rituals no experienced team would spell out for a human.
The four-file repo
A curated index of 1,000+ agent skills from official developer teams and the wider community. The skills are links; they do not live in this repo.
Nobody needs 580 commits to publish a list.
They need 580 commits to keep one true.
The latest commit when I looked was “Fix 15 renamed/transferred repo URLs.” Fifteen upstream projects had moved. Someone found every new address.
The README makes the pitch explicit: “Hand-picked, not AI-slop generated.” Anthropic, Google Labs, Vercel, Stripe, Cloudflare, Trail of Bits and Figma all publish skills. They use different repos, different layouts and no shared registry.
Generating 1,000 plausible skill files is now cheap.
Reading them is not. Testing them is not. Following them when they move is not.
The scarce input moved from writing to attention.
33,000 people starred the receipt.
The code review nobody writes down
mattpocock/skills242k stars37 procedures
The .agents directory Matt Pocock uses with coding agents: reusable instructions for code review, debugging, TDD, design and planning.
code-review/SKILL.mdcodebase-design/SKILL.mddiagnosing-bugs/SKILL.mdtdd/SKILL.mdgrilling/SKILL.md
Its code-review skill does not say “review the code.”
It fixes the comparison point. It resolves the diff against the merge-base. It splits the review into Standards and Spec, then sends each pass to a separate sub-agent so one does not steer the other.
It even supplies a Fowler smell baseline when the repo has no standard of its own.
No senior reviewer needs that script. Giving them one would feel patronizing.
For an agent, the script is the interface. It will follow every line, every time. That makes this level of detail worth writing down.
A scientific library made of Markdown
K-Dense-AI/scientific-agent-skills163 skills117 XML schemas
Ready-to-use agent instructions for scientific software and databases across genomics, chemistry, biology, medicine, drug discovery and quantum computing.
Ten years ago, packaging this kind of expertise usually meant writing a client library.
Here, most of it sits in Markdown: what to query, in what order, what can fail and how to parse the result. The 117 XML schemas tell the agent what the replies look like.
The queries still run as code. What gets reused — the sequence, caveats and output shape — is prose.
9,988 translated files
Microsoft’s beginner course for building AI agents, with written lessons and runnable samples on design patterns, tool use, planning, memory and production.
The English course is about 238 files.
The other 98% is the same material in 54 languages, produced by Microsoft’s own AI translation pipeline.
Making the course available in 54 languages is genuinely useful. Microsoft is also candid about the bulk: the README tells you to use sparse checkout if you do not want half a gigabyte of translations.
Now put that repo beside the four-file index.
The translation pipeline creates more material. The four-file index exists because there is now too much material.
Two edge cases
ai-engineering-from-scratch
A start-to-finish AI engineering curriculum, from math and ML foundations through LLMs, RAG, agents, MCP, evaluation and deployment.
Twenty phases. Five hundred JSON skill definitions. 51k stars — and only six months old when measured.
Structure is valuable in a fast-moving field. It is also the first thing to age.
awesome-gpt-image-2
A gallery of GPT-Image-2 prompts paired with generated results, plus templates and a small application for browsing them.
554 images. Twelve Markdown files. An API directory and a Supabase config.
A prompt library for images has to show its work. Text alone is not enough.
Before agents, a README was enough
public-apis/public-apis473k starsmaintained since 2016
A community-maintained directory of free public APIs, organized by category and published mostly as one large README.
One of GitHub’s most-starred repositories is a page.
That was enough because its reader was a person. A person could scan a table, click a link and recover when the format changed.
The newer repos cannot assume that. They put one capability in each folder, instructions inside the file and machine-readable definitions next door. The useful part has to survive without a browser.
Stars follow distribution
Why did these particular repos break out?
Mostly distribution.
Averaged over its short life, mattpocock/skills gained 1,161 stars a day. The other new repos in this group averaged between 47 and 310 stars a day. public-apis averaged 124 a day for a decade.
Matt Pocock is a full-time TypeScript educator with 42,000 GitHub followers and a former Vercel developer advocate. Two other repos come from companies selling the category they document. Another comes from a developer advocate.
None of that discredits the work.
It does change what the number means.
In this sample, “what is inside” ranges from four files to more than 10,000.
Stars cannot resolve that difference for you.
That still requires judgment.
One repo from the wave's second month gets the same treatment in Humanizer, tested.
Every number came from the GitHub REST API on August 30, 2026 and will drift. Disclosure: I build StashBase, which lets your agent search a local knowledge base. That is why I pay attention to the ones on GitHub — and why I keep a running list of the ones worth knowing about.