Induction pipeline

When you stop recording, ClawMobile runs the trace-induction pipeline to turn your demonstration into a structured skill. It is a record → generate → promote → reuse loop.

Stages

1. Summarize. The raw trace (touches, screenshots, app state) is parsed and a candidate skill is prepared — a draft step graph with inferred selectors.

2. Promote. You review and promote the candidate. It’s written to your workspace skills/<name>/ as version 1 and indexed.

3. Generalize. Concrete values become parameters; steps gain robust selectors and, where needed, branches. This is the generalized_skill.json.

4. Reuse (fast path). On the next run, the skill executes deterministically: each step is matched against the live UI and replayed without calling the model — fast and repeatable.

5. Repair & evolve. If a fast-path step fails (the app changed, a selector no longer matches), the agent reflects on the failure and either repairs the step or hands back to the model. Successful repairs can evolve the skill into a new version.

record ──▶ summarize ──▶ promote ──▶ generalize ──▶ reuse
                                          ▲            │
                                          └── evolve ◀─┘ (on repair)

Fast path vs. agent fallback

The fast path is deterministic: proven steps replay exactly. It fails closed — when a step can’t be matched with confidence, it returns a structured failure rather than guessing, and control returns to the agent. That boundary is what keeps replay both fast and safe.

Generated skills are preview-quality

Induction produces a strong first draft, not a guarantee. Review a new skill’s first few runs, and use the skill drawer to watch its success rate before relying on it unattended.

Closed-loop research

ClawMobile is a research platform for demonstration-driven induction. Beyond metadata updates, structural mutation (changing steps, parameters, and branches across feedback iterations) is an active area — see the project roadmap.

Next: Tools reference