performance

Start by measuring the slow path or high-cost loop, then classify the bottleneck as I/O, build tooling, context size, or algorithmic work. Only optimize after identifying which category dominates, and verify behavior stays unchanged.

Keep context surfaces small and relevant by removing broad repository scans, redundant tool calls, and oversized payloads. Prefer targeted file reads and focused validation over full-suite reruns when it is safe to do so.

Avoid micro-optimizations when you have no measurement evidence showing they address the actual bottleneck. Prefer changes that improve both clarity and speed, and remove unnecessary work before optimizing any code.

Yes, targeted validation is preferred over full-suite reruns when it is safe. Run only the tests covering the changed path, but confirm the optimization did not alter behavior before relying on the reduced scope.

A checklist guides investigation but does not replace profiling tools or benchmarks. It works best for workflow and context overhead; deep algorithmic or systems-level issues still require dedicated measurement tooling.