Most skill directories copy a skill's description and stop there. We install the skill and give Claude a real job to do, then show you what came out.
How each review works
- Install it the way you would. In a new, empty project folder we run the exact install command shown on the skill's page,
npx skills add <owner>/<repo> --skill <name>. If the install fails, the review says so. - Give Claude a realistic task. We write one job of the kind the skill is meant for, phrased the way a user would type it: build a landing page, turn a CSV into a workbook, merge two PDFs. The task is shown word for word on every review.
- Let Claude decide. We do not tell Claude to use the skill. Claude Code chooses skills from their descriptions, so a review also tells you whether the skill gets picked up on its own. We record which skill was actually invoked.
- Run the same task without the skill. A second run in a separate empty folder uses the same prompt with skills switched off. Putting the two results next to each other shows what the skill changes, which is often less, or different, than its description suggests.
- Keep the evidence. We save the full transcript, every file Claude produced, the time taken and the API cost, then screenshot the outputs: web pages in a browser at desktop and phone width, and Word, Excel, PowerPoint and PDF files rendered with LibreOffice.
The test setup
- Claude Code in headless mode, run from the command line, version noted on each review.
- Claude may read and write files and run shell commands, but only inside the test folder, and with no access to our own accounts or keys.
- No MCP servers or connectors, and no skills other than the one under test.
- One run each, with and without the skill. Model output varies from run to run, so treat a single comparison as an example, not a benchmark.
Skills that need an account
Some skills drive a paid or signed-in service: a cloud provider, a payments API, a workplace app. We test the parts that work without an account, such as local development commands or writing configuration, and say clearly on the review what we could not test.
What we do not do
- We do not accept payment for reviews or for placement.
- We do not edit a skill before testing it.
- We do not hide bad results. If a skill did not trigger, failed to install or made the output worse, the review says so.
If you are a skill author and think a review is wrong or out of date, email contact@skillsbot.org and we will re-test it.