How we check
Every guide, comparison and field note on TinkerHood that says something about a skill, tool or workflow shows four checks. Each one is marked checked, not checked or failed, with a date and the version we used. Passing one check tells you nothing about the next, so we report them separately.
The four checks
Format
The files are valid against the published specification. For agent skills, that means the SKILL.md frontmatter and folder layout follow the Agent Skills specification, checked against the specification and with a public validator where one exists.
What it proves: the files are well formed.
What it doesn't: that anything will install, load or help.
Installs
It installs by the documented route, into a clean project, and the files land where the documentation says they should. When there are several routes (copying a folder by hand, npx skills add, a plugin marketplace), we name the route we used.
What it proves: the files arrive in the right place.
What it doesn't: that the agent will find or use them.
Loads
The host actually discovers and uses it. For a skill, we look at the agent's tool trace for the moment it reads the skill's files, rather than trusting an answer that merely sounds right. We check each host separately (Claude Code, Codex, Cursor and others) and list the exact version.
What it proves: this host, at this version, picked it up for a matching request.
What it doesn't: that it improved the result, or that another host behaves the same way.
Helps
It measurably improves the task compared with doing the same task without it. This is the hardest check and the one most often skipped elsewhere. We only mark it checked when we have run a real with-and-without comparison and published the prompts, the scoring and the results.
What it proves: an improvement on the tasks we tried.
What it doesn't: that it will help on yours.
What the marks mean
| Mark | Meaning |
|---|---|
| Checked | We ran it, on the date and version shown. |
| Not checked | We didn’t run it. The page says why (not installed, not measured yet, out of scope). |
| Failed | We ran it and it didn’t work. We say how it failed. |
We never imply a check. If a page is silent about a host, treat that host as not checked.
How we keep it honest
Small samples are labeled: when a result comes from three prompts, we say three prompts. Dates matter, since agent tools change weekly; each page shows when it was last checked, and we re-check when a host ships a change that affects it. Where we can, we publish the exact files and commands so you can run the same check yourself. When we get something wrong, we fix it and note the change on the page. Tell us at hello@tinkerhood.com.
Who checks
We use AI tools in our research and testing, the same tools we write about. The writer named on each page reviews the evidence, edits the page and is responsible for it. Meet the writers.