Prompt cookbook

Use the Stet skill to A/B test my proposed AGENTS.md change. Keep the model
fixed, plan before launch, and return the canonical Trial Result with its
recommendation, evidence limitations, and next action.

Use the Stet skill to improve AGENTS.md within a bounded search space and stop
rule. Test one change at a time and ask before using holdout evidence or
promoting a finalist.

Use the Stet skill to test whether this repo-managed skill helps. Plan before
launch, use the right baseline, and report promote, hold, or inspect.

Use the Stet skill to compare these configurations on the same repository
tasks. First ask whether I want a cheap diagnostic read or a gateable rollout
decision. Keep diagnostic evidence labeled directional; for a gateable choice,
use matched runs and return the canonical Trial Result.

Use the Stet skill. Onboard this repo for Stet evals. Ask what work I want
Stet to track, read CI first, and choose the real test command. Inspect merged
PRs/commits for representative work (use read-only subagents for this
investigation when available), create the Harbor setup, and build a
representative starter dataset. Choose and pass the narrowest credible bounded
verifier to `stet suite build` with `--test` (alongside the manifest and a
bounded `--workers` value). Treat `bazel test //...`, unfiltered `pytest`,
`go test ./...`, and workspace-wide scripts as broad; do not launch one without
my explicit approval. If no credible bounded verifier exists, stop before
building and propose alternatives. Do not broaden merely to increase yield.
Report scope, skipped tasks and reasons, then stop before any model eval. Treat
Docker or replay checks as setup validation, not a model eval.
Wait for `build-summary.json` before treating the build as complete, then return
the onboarding receipt with the selected slice, scope, skips, and confidence.

Use the Stet skill. Expand the existing onboarding slice by one bounded
subsystem or task family. Read the current receipt first, preserve its test
command with an explicit `--test`, and explain the expected coverage gain.
Use read-only investigation where useful; do not broaden just to increase
yield. Stop before building if no narrow verifier is credible, and propose
alternatives. Report added scope and skips, then stop before any model eval.

Use the Stet skill. Check whether this repo is ready for Stet. Inspect CI,
build files, test commands, Docker assumptions, and model-provider auth. Report
blockers before editing files or launching evals.

Use the Stet skill. Read the current Stet result from status/report surfaces.
Tell me the recommendation, confidence, evidence quality, grader coverage,
task coverage, next action, and residual risk. Do not reconstruct the verdict
from pass rate alone.

Use the Stet skill. Check the current Stet run. Use status/report surfaces
only, tell me whether it is running, stalled, exited, repairable, or complete,
and give the exact next command.

Use the Stet skill. Keep the workflow bounded. Read status/report artifacts as
the source of truth, change one lever at a time, treat inspect as a diagnostic
state, and ask before launching expensive evals or discarding existing
evidence.