Thanks for wanting to add to this. The bar is simple: a skill has to do one job well and be usable by a stranger.
Each skill is a folder with a SKILL.md at its root:
skills/<skill-name>/
├── SKILL.md # required — frontmatter + instructions
└── reference/ # optional — catalogs, examples, configs Claude loads on demand
SKILL.md frontmatter needs a name and a description. The description is what Claude uses to decide when to trigger the skill, so make it concrete: say what it does and list the trigger phrases.
---
name: my-skill
description: What it does, in one or two sentences, plus when to use it. Trigger on "phrase," "phrase," "phrase."
---
# My Skill
Instructions here. Keep the SKILL.md short; push detail into reference/ files.- Put the skill under
plugins/<plugin-name>/skills/<skill-name>/. - Add
plugins/<plugin-name>/.claude-plugin/plugin.json. - Add an entry to
.claude-plugin/marketplace.json.
- Depersonalized. No private names, tokens, or company-only rules baked in. Make house rules configurable instead.
- Sourced. If a skill makes factual or research claims, cite them.
- Tested. Show that it works — a before/after, a sample run, something.
- Tight. One job. If it's doing three things, it's three skills.
By contributing you agree your work ships under the repo's MIT license.