The Manifest · Issue 002 ·
The number your agent trusts most is the one nobody can audit
Our most-visited page tells your agent to check it first — before it even searches.
Hi again.
This week: an install count nobody can audit, six additions to the directory, what 199 MCP servers revealed under inspection, and a workflow that turns a Figma file into React.
THE SIGNAL / 01
Before your agent searches for a skill, it’s told to sort them by a number nobody outside the registry can audit.
find-skills by Vercel Labs is the most-visited page on AgenticSkills. Between August 14 and 26, it accounted for 46% of our organic search clicks — up from 5% the month before.
It’s a good skill. It does something genuinely useful: your agent searches the registry instead of guessing at package names.
But before it searches, Step 2 tells the agent to “Check the Leaderboard First”:
“The leaderboard ranks skills by total installs, surfacing the most popular and battle-tested options.”
Most popular is a measurement. Battle-tested is a judgement — and an install counter can’t tell you whether software holds up in production. A skill pulled ten thousand times by CI runners and abandoned in week two can outrank one that fifty engineers depend on daily.
Then Step 4 tells the agent to prefer skills with 1K+ installs and be cautious below 100. Step 5 hands that same number to the user as a credential.
So install count influences what the agent looks at, what survives, and what gets recommended — all resting on a number only the registry reporting it can see.
We used to display install counts too. We dropped them because we couldn’t defend them.
Instead, we show what can be checked against the source: repo stars, the named author, contributors to that specific path, and the date of the last commit to the skill’s own directory. A skill inside a 40K-star repo that hasn’t been touched in fourteen months shouldn’t inherit the repo’s credibility.
None of those signals are perfect. All of them are verifiable, which is the whole difference.
NEW & VETTED / 02
Two new, and four we owe you.
Last week we said six design skills landed. Ten did. Here are the four we didn’t name, plus the two that cleared since:
Two of those are worth a sentence. Magic Hour is our first MCP server listed by submission rather than by us — the pipeline that receives them had been broken, silently, and 001 went out while it still was. And the miscount above is ours. Ten skills landed on August 27; we told you six. Nobody wrote in. We publish the correction anyway, because an issue that spends its first section on unverifiable numbers doesn’t get to be careless with its own.
THE PICK / 03
Impeccable
S-RANK · 65.8K+ STARS · 46 CONTRIBUTORS · AS OF SEPTEMBER 5, 2026 · BY PAUL BAKAUS
A design language that pushes an agent past safe, generic UI.
Most design skills hand the agent a style guide. This one makes it commit to a position first. Every surface gets one of four modes — Persuade, Operate, Read, Experience — chosen from the surface, not the product. A developer tool’s landing page is still Persuade. A fashion house’s documentation is still Read. The mode decides which side wins when expression and scannability conflict, which is the argument that usually goes unresolved and produces the mush.
The part we didn’t expect: it puts a ceiling on its own QA.
“Build fully, inspect once with a batched round […], fix everything it shows in one batch, confirm with at most one more round, and stop polishing.”
The file says open-ended self-QA “burns the user’s money.” That’s a skill author choosing to spend fewer of your tokens, written into the skill. We’ve read a lot of these. It’s rare.
It’s also not a plain SKILL.md. It ships a CLI, a plugin, hooks, and agents, and generates output for every major agent runtime from one authored source — which is why it installs with npx impeccable install and not npx skills add.
And the caveat we’d want to know: 65,000-plus stars, and one person has written 1,328 of the commits. The next human contributor has 146. The count on the path we index reads 3, because that path is generated output; authorship lives elsewhere in the repo. Nothing here is wrong. But by the standard in 01, a big number attached to a narrow point of failure is exactly the kind of signal you should look past rather than lean on. We listed it because the source is legible, not because the star count is large.
$ npx impeccable install
DIDN’T PASS / 04
199 MCP servers, eight automated checks
We audited all 199 MCP servers in the directory against eight automated checks. Of those, 86 had no public source, 80 were hosted endpoints, and 33 had public repositories. Thirty-one of those repositories could be evaluated against our repo-level checks.
- 86 of 199publish no auditable source at all. Not a failed check — nothing to check. Just under half the catalogue.
- 31 of 31auditable repositories publish no SBOM. Not most. All of them.
- 70 of 80hosted servers expose no RFC 9728 OAuth metadata, so a client can’t discover how to authenticate without out-of-band documentation.
- 10 of 31repositories have no SECURITY.md — no stated way to report a vulnerability privately.
The good news is narrow but real: every audited repo carries a license, and no hosted endpoint failed TLS.
We apply the same thing to ourselves, and it costs us entries. Last catalogue sweep removed twelve skills. Some were dead repos. One, testing-anti-patterns, was attributed to a real project’s repository and simply wasn’t in their tree — a skill that didn’t exist, listed by us, until someone checked. We publish that number because a directory that never removes anything isn’t curating.
WORKFLOW / 05
Design-to-Code
Claude reads your actual Figma file and generates React that matches it. Intermediate · about 10 minutes to set up.
Three skills — frontend-design, figma-implement-design, react-best-practices — wired to one MCP server: Figma.
It’s the natural counterweight to 03. Impeccable is for when nobody has decided what the thing should look like, and the agent has to commit to a position. Design-to-Code is for when somebody already did, and the job is to not lose it in translation. Reach for the wrong one and you get an agent improvising over a design system that already had the answer.
There’s also a quieter reason it belongs in this issue. A Figma file is the rare case where the design brief is machine-readable: components, variables, spacing and colour are facts the agent can read rather than infer. When a checkable source exists, use it. Don’t let the model guess at something it could have looked up.
That’s it for Issue 002. See you next week.
— The Manifest