What problem does it solve? Improving an existing agent skill often turns into open-ended, unmeasured tweaking. This Skill enforces a bounded autoresearch loop with one primary metric, a fixed 3-5 scenario harness, and explicit keep-or-reject decisions so skill quality ratchets upward without uncontrolled experimentation. ## Core Features & Use Cases - Workspace Scaffolding: A bash script creates a durable autoresearch workspace under ~/.AGENTS-temp/<repo>/autoresearch/<skill>/ with goal, harness, baseline, iteration template, and decision log files. - Metric and Harness Guidance: A reference document defines good primary metrics (first-try success rate, retries, manual commands) and harness shapes for operator, build-and-validate, and investigation workflows. - Keep-or-Reject Iteration Log: Each iteration makes one bounded change, re-runs the same harness, and records a keep, reject, or defer decision. - Use Case: You have a skill like imagebuilder-bake-validate that works but requires too many manual commands. Run the scaffold script, define "first-try success rate" as the metric, record a baseline, then iterate one change at a time until the metric improves. ## Quick Start Ask the agent to scaffold an autoresearch workspace for a chosen skill and repo with a primary metric, then record a baseline before making any changes.
