Claude Code skills

Versioned, model-invoked capability modules that extend Claude Code with domain procedures, loaded on demand rather than held in the base prompt.

How it works

Each skill is a self-contained module with frontmatter metadata describing when it applies. The model scans those descriptions, then loads a skill on demand for a matching task signal, so the procedure lives in the module rather than in the always-on base prompt. Routing happens on the description rather than on conversation context, which is how the model decides to reach the procedure at all.

Why it matters

Routing by description keeps the base prompt small while reaching deep capability for tasks the description anticipates. The trade-off is that the routing depends on the description matching a task signal, so a task the user cannot articulate in those terms never triggers the skill, and a sloppily-worded description either fails to load or loads for the wrong task. Description-tuning becomes its own surface, competing with prompt engineering.

In practice

A task with the right wording, such as an explicit request to write a doc, triggers a writing skill whose description names that case; an adjacent task phrased without those signals leaves the skill unloaded, even though the procedure would apply.

Defined by Ready Solutions AI