Sources over
superlatives.

Read the source. Keep the judgment separate.

Entries are based on project repositories, skill instructions, documentation, and licenses. Each review links back to the original material. The verdict, best-fit guidance, pairings, and cautions are editorial interpretations of those instructions.

They are not the result of a controlled benchmark. We have not published measured design-quality, accessibility, speed, or conversion scores. A recommendation is a starting point for your own project, not a guarantee of an outcome.

Unknown is a useful answer.

Agent support is marked native when the source documents support, partial when setup or adaptation is needed, and not verified when the available evidence is insufficient. A missing compatibility claim does not mean a skill cannot work with an agent.

Scripts and hooks have three states: present, none found in the inspected skill, and not verified. These describe the reviewed contents. They are not a security audit. Missing licenses and install commands stay missing until they can be established from the source.

Popularity is context, not a verdict.

Stars, forks, open issues and pull requests, and the last commit come from GitHub repository snapshots. Several skills can share one repository, so these are repository totals, not individual skill adoption. The snapshot date appears with the counts; snapshots older than 72 hours are flagged on detail pages.

If a refresh fails, the last successful snapshot keeps its original date. If no snapshot exists, the UI says so. Missing values do not become zeros. Installation counts are omitted when a reliable source is unavailable.

Install with a clear scope.

Install controls display curated commands as text. Agent-specific instructions take priority when available. A universal command can still need agent selection or further setup; it does not promise that your chosen agent is configured automatically.

Review a project’s current instructions and the files it will add before running a command. Sources can change after curation. The Stack Builder keeps unsupported command formats separate and surfaces overlapping roles and recorded conflicts.

What comes next.

Reproducible benchmarks and a DESIGN.md guide are planned for a later edition. Community submissions are not open yet. Those unfinished sections stay out of search indexing until they contain published work.