What problem does it solve? Caveman optimization reports contain aggregate observations that are easy to misread as savings estimates or ready-made fixes. This Skill enforces an evidence-first workflow that turns a report-only observation into an operator-approved candidate change validated by a paired baseline evaluation, preventing unsupported claims about cost savings or code safety. ## Core Features & Use Cases - Observation Reading: Runs caveman opportunities list and reads only the report_only_observations array, preserving server-provided titles and observations verbatim while rejecting retired profile ids. - Operator-Chosen Selection: Presents supported observations without ranking and requires explicit operator approval before inspecting callsites or editing code. - Paired Baseline Evaluation: Designs and runs baseline versus candidate measurements on identical fixed inputs, recording quality checks, token or byte costs, fixtures, and confounders before keeping or reverting a change. - Use Case: A Caveman report shows a tool-output-size-profile observation. Use this Skill to present it to the operator, design one minimal candidate change at the evidenced callsite, run a paired eval on a fixed fixture, and report keep, reject, or inconclusive without claiming dollar savings. ## Quick Start Ask the assistant to evaluate the current Caveman optimization report and walk through the supported observations for an operator-chosen candidate with a paired baseline evaluation.
