Every other directory counts its inventory. We ran ours. Each skill got a realistic task, then the identical task was run again with the skill removed, scored against a fixture with a known set of planted defects. Five of the six did not change the output enough to justify installing them, so they are not here. The results table below shows all six, including the failures.
The uncomfortable result first. Five of the six skills we wrote did not beat a competent assistant working without them. These were not lazy skills. Each is roughly 2,500 words of real methodology, and we checked the existing directories first so we were not rebuilding something that already existed. They still failed.
Here is every result, unedited.
| Skill | The test | Without it | With it | Verdict |
|---|---|---|---|---|
| Interface state audit | Build a team-members admin screen. Score 18 interface states drawn from real failures. | 14 of 18 | 18 of 18 | Published |
| Blast radius review | Review a fix for an access-control bug. 25 files, 9 directories, 5 more instances of the same bug scattered across routes, a cron job, a GraphQL resolver and a websocket handler, plus 5 decoys that only look wrong. | 5 of 5 found, 5 of 5 decoys correctly cleared | Identical | Cut |
| Contract drift review | Review a new endpoint whose code names 6 columns, tables and routes that do not exist. | 6 of 6 | 6 of 6 | Cut |
| Design system conformance | Add a toast component to a codebase with an existing token system. | 0 hardcoded colours, reused the existing button, guarded reduced motion | Same, minus 2 off-scale pixel values | Cut |
| SERP-first content brief | Plan content for 3 keywords where the obvious answer is wrong for 2 of them. | Reached all 3 correct format calls, including "do not write an article, build a tool" | Same calls, plus a two-page split on one keyword | Cut |
| Pre-publish page audit | Audit a page carrying 13 planted defects, one of which makes the other 12 irrelevant. | 12 of 13, and led with the right one | 13 of 13 | Cut |
The one skill that worked supplies knowledge. The five that failed supply discipline.
A modern assistant already searches thoroughly, already checks identifiers against the schema when the schema is in front of it, and already knows to lead an audit with the finding that makes the others moot. Writing that down in a file and asking it to follow the file changes very little, because it was going to do that anyway. That is a real and slightly deflating finding about the entire category, and it is why directories advertising two million skills are selling something closer to decoration than to capability.
What an assistant does not do reliably is enumerate things it was not prompted to think of. Ask for a members screen and you get a good members screen: loading, errors, empty search, seat limits, long names. You do not get the first-run empty state, the partial failure when three of five invites succeed, the guard that stops you demoting the last administrator, or the focus placement after a row is deleted. Those are not failures of thoroughness. They are gaps in a checklist nobody was holding. Hand over the checklist and all four appear.
So the skills worth installing are the ones carrying a taxonomy, a domain checklist, or a set of facts the model will not generate unprompted. The ones telling a capable assistant to be careful are telling it something it already knew.
- It has to be a method, not a mood. If it could be replaced by one sentence in a system prompt, it is a sentence. - It has to beat the control. Same task, run twice, scored against planted defects. Every page shows the number, including where the gap was small. - It has to be different from what already exists. We read the official skills and the large community collections first. - It has to admit what it cannot do. Every page carries a plain list of what the skill does not cover.
Nothing here needs an account or our app. Copy the file or download it and it works today. Locul just makes it one click and puts the file where each assistant expects it.
Skills are organised by the job you are doing, not the job title you have.
Locul puts a skill in the right folder for Claude Code, Claude Desktop and Cursor, scrubs any environment variables before they reach you, and keeps the memory those skills read from up to date by itself. Mac and Windows.