What are skills?
A skill is a reusable procedure the agent learns by watching you do something once. Instead of re-prompting the model every time, ClawMobile records your demonstration and turns it into a structured, replayable artifact.
Why skills
Re-driving an app from natural language on every run is slow, costly, and non-deterministic. A skill captures the proven steps so subsequent runs:
- replay on a deterministic fast path when the UI matches expectations,
- fall back to the agent only when something has changed,
- and improve over time as later runs feed back into the skill.
Anatomy of a skill
Each skill lives under your workspace skills/<name>/ and is mirrored to R2:
| File | Purpose |
|---|---|
SKILL.md | Human-readable description and trigger |
generalized_skill.json | The induced step graph (selectors, params, branches) |
trace/ | The source recording (screenshots, getevent, app/window state) |
Versioning
Skills are versioned. Every change — a re-recording, a metadata tweak, or a
structural mutation — creates a new version in R2 (<user>/<skill>/v<n>/) and a
row in the skill index. The skill drawer shows
the full version timeline and recent evaluation runs, so you can see how a skill
evolved and how reliably it runs.
| Mutation type | What changed |
|---|---|
initial | First induction from a demo |
metadata | Description / trigger / parameter labels |
structural | Steps, parameters, or branches (research feature) |
Sharing
Any skill can be shared via a time-limited, presigned link. The recipient gets a read-only copy of the generalized skill — never your recording or credentials.
Next: Recording a demo