How to publish an agent Skill and get it listed

These are the things that most often stop an otherwise good Skill from being usable — or listable.

A stable, fetchable path

SKILL.md at a path that will not move: typically skills/<name>/SKILL.md in a public repo. If the path changes, every install command and every citation breaks.

A real LICENSE file

Not a README badge, not a frontmatter field alone. We have verified repos whose badge claims a licence the LICENSE file does not, and repos with no licence file at all — which legally means all rights reserved. This is the single most common gap we find.

A description that triggers

The largest determinant of whether your Skill ever runs. Name the artefacts, the phrasings, and the situations. The strongest vendor Skills also state when NOT to use them and which sibling Skill to use instead.

A one-command install

Document the exact command per agent. A clone-and-copy install is a real friction cost and scores lower on ergonomics than a one-liner, on this site and in practice.

Reference files split by case

Keep SKILL.md under ~5K tokens and put the detail in reference files, one per symptom or task. Cheaper to trigger and easier to maintain.

Bundle scripts where determinism matters

A tested script beats instructions telling the agent to write the same code again. It is also what separates a Skill from a long prompt.

Every entry is verified against primary sources, not a crawl. We fetch the repo, confirm the SKILL.md exists at the exact path we will publish, read the LICENSE file, record the real last-push date, and test the install command. The date on the entry is the date that happened.

We also classify provenance honestly: Anthropic, verified organisation (a real company publishing for its own product), or community. And we list only per-agent compatibility we have actually tested — a Skill that plausibly works on five agents gets listed for the ones we ran it on.

The bar is editorial as well as factual. We do not add a Skill because it exists; we add it because a developer browsing that category would expect to find it. Category depth is the goal, not catalogue size.

Open an issue or pull request on the site’s repository with the repo URL, the exact SKILL.md path, the licence, and the install command for each agent you have tested. If you have install or usage numbers, include the source.

The strongest submissions do our verification for us and are honest about limits: what the Skill is not good at, which agents you have not tested, what credentials it needs. A stated limitation makes an entry more useful and more likely to be listed, not less — the notIdealFor field on every entry exists for exactly that.

Entries are also removed. If a repo is archived, a licence changes in a hostile direction, or an install command stops working, the entry goes and a permanent redirect is left in its place. Listing is a maintained claim, not a one-time favour.