What problem does it solve? Editing a skill, AGENTS.md, or system prompt often feels like an improvement but there is no proof. This Skill turns guidance changes into controlled experiments: pin the task, change one variable, run fresh agents before and after, and compare tokens, tool calls, failures, and correctness so you keep only changes that measurably help. ## Core Features & Use Cases - Controlled A/B methodology: Freeze prompt bytes, product revision, model, reasoning level, auth, and backend state so only the tested axis varies between runs. - Structured evidence capture: Record wall time, input/output/reasoning/cache tokens, model tool calls, surface actions, failed actions, retries, outcome, and cleanup per trial in a defined metrics table. - Decision framework: Rank results by correctness first, then policy compliance, cleanup, failures, wall time, tool calls, and tokens, with a standard report format. - Use Case: You rewrote a routing section in AGENTS.md. Run three fresh-agent trials of the same read-only task before and after the edit, then compare failed actions and input tokens to decide whether to keep the change. ## Quick Start Ask the agent to benchmark whether a recent edit to AGENTS.md reduced tool calls and tokens by running the same fixed prompt with fresh agents before and after the change.
