Prompt cookbook

Use the Stet skill to plan a comparison of this new arm on the existing corpus.
Reuse valid current evidence and spend only on the new arm. State the frozen
baseline, matched slice, verifier, cost boundary, and plan before doing spend.

Use the Stet skill to inspect the freshest existing Trial Results and compare
artifacts. Do not launch models. Tell me what can be reanalyzed without spend,
what remains uncertain, and the one next action if new evidence is necessary.

Use the Stet skill to propose one bounded improvement to this instruction file
or repository-managed skill. Name the baseline, one lever, retained slice,
verifier, stop rules, and spend boundary. Confirm the plan before model spend.

Use the Stet skill to inspect this run's status and next_action. Classify the
state, say whether recovery spends model tokens, and execute only an approved
evidence-preserving action. Do not restart a healthy wait or discard bad rows.

Read the canonical Trial Result on my behalf. Report the routing verdict,
confidence, denominator, per-metric calibration, validity, confounds, observed
spend, claim limits, and one next action. Do not reconstruct from pass rate alone.