Skip to main content
extract-skill looks back at a run (or a pattern you describe) and harvests the durable, reusable part of it, turning a one-off into an agent skill doc, a sketch for a new workflow, and memory facts worth keeping. Reach for it right after a run when you notice yourself doing the same thing again.
bunx smithers-orchestrator workflow run extract-skill \
  --input '{"targetRunId":"RUN_ID","prompt":"Harvest the way we triage flaky tests into a reusable skill"}'
Pass targetRunId to let the analysis read the real run state (bunx smithers-orchestrator inspect, events, node); leave it null to harvest purely from the prompt and context. When the pattern is skill-shaped, the workflow writes the skill to .smithers/skills/<name>.md.

Stages

  1. analyze: read the run (if given) and decide what is worth harvesting: the repeated pattern, whether it is reusable as a skill and/or a whole workflow, and any durable memory facts.
  2. propose: turn the analysis into concrete proposals: a skill (name, description, body), a workflow sketch (id, sketch), and the polished memory facts to persist.
  3. scaffold-skill: when the pattern is reusable as a skill, write the proposed skill doc to .smithers/skills/<name>.md.

Inputs

InputTypeDefault
targetRunIdstring or nullnull (analyze the prompt/context alone)
promptstring"Describe the pattern or run you want to harvest into a reusable skill, workflow, or memory."

Monitor it

bunx smithers-orchestrator workflow run extract-skill --input '{"targetRunId":"RUN_ID","prompt":"…"}' -d   # detach
bunx smithers-orchestrator ps                                                                       # active / paused / recent
bunx smithers-orchestrator logs RUN_ID -f                                                          # follow events
bunx smithers-orchestrator inspect RUN_ID                                                          # full run state
Use this to close the loop after a run: instead of re-deriving the same approach next time, harvest it once into a skill, a workflow sketch, or memory the next run can lean on.