Choose a workflow

Each workflow changes one meaningful harness surface at a time, uses a matched task slice, and returns a Trial Result you can interpret before deciding what to keep. Choose the question that matches the decision in front of you.

Is my AGENTS.md change helping?

Use a manifest-backed A/B comparison when you have a proposed instruction change and want the model and other relevant settings held fixed. Recommended prompt

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.

Evidence expectation: A credible result compares baseline and candidate on the same retained repository tasks, with the intended instruction lever isolated and the required tests and graders present. A quick smoke can guide iteration but is not a rollout decision.

Improve my AGENTS.md

Use the iterative loop when you want the agent to propose bounded instruction edits and measure them one at a time. Set a search space and stop rule before launching; holdout evidence and promotion remain explicit decisions. Recommended prompt

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.

Evidence expectation: Each iteration should name the single changed lever, preserve the task and grader contract, and stop when the result is inspect-only, not improving, or otherwise unable to support the next decision.

Is this skill helping?

Evaluate a skill against the right baseline: an effectively absent skill for a new skill, or the committed skill for a revision. Keep repository work and behavior-relevant graders matched. Recommended prompt

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.

Evidence expectation: The baseline must actually represent skill absence or the committed version, and the result should include behavior-relevant grader coverage. Missing coverage turns a comparison into an inspectable diagnostic, not a keep decision.

Which model or reasoning effort should I use?

Compare models on the same retained tasks, or hold the model fixed while comparing reasoning settings. Keep the rest of the harness stable enough to attribute the observed difference. Recommended prompt

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.

Evidence expectation: Read correctness, quality dimensions, cost, uncertainty, validity, and residual risk separately. Do not choose a default from stale, partial, incomparable, or inspect-only evidence.

What happens next

Start with the prompt cookbook for compact copy-paste wording. When your run finishes, read the Trial Result before making a rollout decision.

​Is my AGENTS.md change helping?

​Improve my AGENTS.md

​Is this skill helping?

​Which model or reasoning effort should I use?

​What happens next

Is my AGENTS.md change helping?

Improve my AGENTS.md

Is this skill helping?

Which model or reasoning effort should I use?

What happens next