The most surprising thing about reading 1,291 agent skill specs in a row is how similar they are. Same shape. Same “Use this when…” trigger pattern. Same three or four examples blocks. Same five-paragraph “Implementation notes” coda. The format has converged faster than the ecosystem itself.
What hasn’t converged is everything around the spec — whether the repo has a license, whether the README has a screencast, whether someone has touched the code this quarter. We score every entry on two independent axes — Craft (is the SKILL.md written well?) and Adoption (is the thing it lives in usable?) — and we publish the ones that clear a bar on both. The arithmetic across the directory is harsh:
- 1,291 scored entries
- 249 publish (composite ≥ 4.0 / 5)
- 337 sit in the borderline tier (5.0 to 7.9 internal)
- 705 never even make curation
That’s a ~19% publish rate. The drop from 1,291 to 249 isn’t a Craft problem. We computed it.
Average Craft across the directory: 2.82 / 5. Average Adoption: 2.16 / 5. Both are mediocre. But the gap between the best and the worst is wider on Adoption than on Craft. SKILL.md files mostly cluster within one point of each other on the rubric. They differ on whether the GitHub repo around them has been touched in the last 90 days, whether the README is more than 500 characters, whether a license file exists.
The five hidden gems
Here’s the most uncomfortable finding. Five SKILL.md files in the directory score Craft ≥ 4.5 / 5 — by our rubric, exceptionally well-written specs — and Adoption ≤ 2.1 / 5. They are:
tock— Craft 4.7, Adoption 2.0audio-tts— Craft 4.5, Adoption 2.1ffmpeg— Craft 4.5, Adoption 2.1aisync— Craft 4.7, Adoption 2.1btw— Craft 4.8, Adoption 2.1
These are brilliantly written agent skills nobody can find. Their triggers are explicit, their outputs are named, their scope is precise. By the SKILL.md format’s own standards, they are model citizens. And the repos around them are stagnant — no license, no examples block, no recent commits. Our composite (harmonic mean of the two axes) drops them out of the top picks. They’re discoverable via search and the categorical pages, but they don’t surface on the homepage.
If you’re the author of one of these and you’re reading this: add a license file. Add a 50-line examples/ directory. Push a commit. The Craft score doesn’t move; the Adoption score jumps a full point; your composite breaks 4.0 and you appear next to the official Anthropic skills on the homepage. It is the cheapest leveraged effort in the ecosystem.
The license disaster
Here’s a number we did not expect.
1,278 of 1,291 scraped entries have no license field in their GitHub metadata.
That’s 99.0%. The thirteen exceptions are mostly the official anthropics/skills and openai/skills catalogs (MIT and Apache-2.0 respectively), the few skillsmp entries that bothered, and a handful of profiled builders like zarazhangrui and swyxio whose major repos ship MIT licenses.
This is downstream of the SKILL.md format’s design. The spec only describes the file inside the folder; it has nothing to say about whether the containing repository is legally usable. So most authors don’t think about it. A skill that doesn’t ship a license is, in most jurisdictions, all-rights-reserved by default — you can read it, but technically you can’t even fork it without exposing yourself.
Our A3 dimension treats no-license as a neutral score (we don’t punish unawareness) rather than a zero. But it’s a soft cap on Adoption that catches almost every entry. If your goal is for actual builders to adopt your work, this is a one-line fix that moves your score.
The leaderboard isn’t where you think
Look at the top six picks on the homepage today, sorted by composite:
- swyxio/spark-joy — 4.5 (Adoption only)
- zarazhangrui/beautiful-html-templates — 4.4
- zarazhangrui/frontend-slides — 4.4
- swyxio/swyxio — 4.3
- swyxio/ai-notes — 4.2
- zarazhangrui/follow-builders — 4.0
None of these are SKILL.md files. They are regular GitHub repositories — the kind we index as repo_tool rather than skill_md — that we surface because their authors are in our editorial builder roster and the work is usable as-is by an agent. They get an Adoption-only score and pass the bar because they have licenses, docs, and recent activity.
The first SKILL.md-format entry on the list is the Anthropic doc-coauthoring skill at composite 4.0 (Craft 4.8, Adoption 3.4). The Craft is a full point higher than spark-joy’s hypothetical Craft would be — but the composite is lower, because the Adoption axis is harsher on small, low-traffic specs than on giant well-loved repos.
This says something about where the agent-skills ecosystem actually lives in 2026: the format is winning the intent — describing trigger-driven tools — but most of the actual adopted tools haven’t been ported to the format yet. The top of the leaderboard is the bridge between the two worlds.
The diegosouzapw question
We index 346 builders. Most have one or two published skills. A handful — Anthropic, OpenAI, the marketplaces — have a dozen or more. And then there is diegosouzapw, an account with no website, no Twitter visible from the profile, and 30 published skills.
The skills are eclectic — status (cli-and-api), dig (personal-system), git-sw (agent-infrastructure), anime, browser, zift, boggle, spacex. Average Craft is solid. Average Adoption is low. It is a one-person research program in public, and we can’t tell you who runs it.
This is the pattern we did not anticipate. The agent-skills ecosystem has a long-tail of single-author shadow catalogs that haven’t been written about, don’t appear in marketplace rankings, and would never have made any list that wasn’t a daily-scrape with a permissive admission policy. Our directory exists partly to give the diegosouzapws of the world a shot at being read. Whether they want to be read is a separate question — but the option is now on the table.
What this means if you’re shipping a skill
Three asymmetries we’d extract from the data:
-
Craft is a one-time cost; Adoption is a sustained cost. Writing the SKILL.md well takes an afternoon. Maintaining the repo around it, picking a license, writing examples, replying to issues — that’s the actual moat. Skills die from neglect, not from bad triggers.
-
The Adoption signal is hackable in five minutes. Add a
LICENSE(MIT, copy-paste). Add anexamples/folder with two real usages. Push one commit per month. These moves cost essentially nothing and they each shift your A1/A2/A3 a notable amount. We’ve watched a borderline-tier skill clear the publish threshold from a single commit that touched the README. -
Where you publish matters less than that you publish anywhere indexable. Our pipeline finds SKILL.md files via GitHub code search and the major marketplaces. If your skill is in any of those — even on a four-star repo with no following — it will be scraped within 24 hours. The bottleneck for entry is publication, not promotion.
Coming up
A few things we’re working on against the data above:
- A public companion index of the top 50 picks that re-syncs daily.
- A growing set of builder profiles — editorial five-block reads of the people doing disproportionate work in the ecosystem. We’re at two (Zara Zhang, Shawn Wang / swyxio); the next ten are queued.
- Adoption metadata backfill for the 1,278 entries currently missing license / commit data, which will move the harmonic mean closer to a true composite across the whole directory.
If you build skills, the most useful thing you can do for the field right now is keep the repo around your spec alive — a license, a recent commit, a working example. The format is doing its job. The repos around it are the slow part.
If we have misread your work, open an issue and we’ll re-score in the next 24 hours.
Have a comment or correction? We’d rather get the score right than be quick. The raw data behind this post is in the source repo — every number reproduces from data/db.sqlite.