find-skills

/find-skills — mine the Rockie skill catalog

The Rockie platform ships ~300 skills. They are deliberately not in your context. Loading 300 descriptions at every session start would cost more than the skills are worth. Instead the catalog sits behind the CLI, and you pull the two or three that match the work in front of you.

That is the whole design: the catalog is your library card, not your bookshelf. This skill is how you use it.

A Rockie SKILL.md uses the same frontmatter Claude Code expects. A pulled skill is invocable in the same session — no restart, no registration step.

When to reach for this

• The task names a framework/library/tool you'd otherwise wing it on (vLLM, verl, Unsloth, TRL, SGLang, RDKit, Biopython, DuckDB, …).
• You're about to write a long block of domain guidance from memory.
• You're entering a domain the project hasn't touched before.
• The user asks "is there a skill for X?"

Check the catalog before writing expert guidance from scratch. A pulled skill is written by someone who has actually run the thing.

Browse

rockie skill catalog --search grpo --json
rockie skill catalog --category ml-inference --json
rockie skill catalog --json                       # everything

Always use --json. Not for parsing convenience — because the tab-delimited form omits a field you need. Shape:

{"skills": [{"name": "verl-rl-training", "catalog_id": "verl",
             "description": "...", "category": "ml-training", "resources": [...]}],
 "catalog_revision": "f138f4b…",
 "tenant_visible_count": 296}

The two identifiers — read this before you pull

name and catalog_id are different fields and both matter. They differ for 39 of 297 catalog entries (~13%):

field	what it is	you use it for
catalog_id	the catalog slug (shown as /slug in tab output)	rockie skill pull <catalog_id>
name	the skill's frontmatter name	invoking it: /<name>