Optimize

Own the search for a measurably better candidate: the experiment question, comparison, and stop decision. A supported no-improvement result is valid. An unknown implementation alone does not require an experiment: use this skill when comparative trials and a measured keep/revert decision are needed to establish the requested result. Keep ordinary implementation or validation with its owner.

Bound the experiment

Establish the target, representative workload, measurable objective, hard constraints, mutable scope, current candidate, available tools, and permitted effects. Use existing project knowledge when it could change the hypotheses or proof.

Set a finite resource budget covering trials and confirmation before starting. Use a small bounded run when none was supplied; never initiate uncapped paid work. Analysis alone does not authorize edits. Installation, new dependencies, external disclosure, publication, and destructive actions retain their own authority requirements.

Establish a valid comparison

Prefer an existing benchmark, profiler, runtime probe, or evaluation surface; record the baseline before changing the target. If required measurement, host access, or credentials are unavailable, report the gap rather than simulate a result.

The metric must represent the requested outcome. Correctness, security, required behaviour, compatibility, and resource limits are hard constraints: an aggregate score cannot compensate for violating them.

Trust the baseline

Use representative inputs and pin the candidate, data, dependency/runtime versions, host configuration, and permissions that materially affect the result. Check that the harness exercises the target and propagates failed or missing results. Choose the useful improvement threshold and comparison method before seeing candidate measurements.