Never optimize on a hunch. Take a measurement — a profile, a bundle analysis, a Core Web Vitals reading — locate the actual bottleneck, then fix that and re-measure to confirm the win. Findings without measurement are labeled potential impact, never presented as fact (the metric-honesty rule). Prevents speculative complexity that trades readability for imaginary performance.
Applied by (backlinks)
Addy Osmani — Agent Skills:
- addy-performance — Core Web Vitals targets, profiling workflows, bundle analysis.
- addy-web-performance-auditor — Quick mode (“potential impact”) vs Deep mode (measured, from Lighthouse/PSI/CrUX/trace data).
Compound Engineering:
- ce-optimize — baseline → parallel experiments → three-tier evaluation; never optimize without a metric.
- ce-performance-oracle — bottleneck analysis backed by measurements; ce-repo-research-analyst — finds the metrics/benchmark surfaces first.
gstack:
- gstack-benchmark — baseline Core Web Vitals / load, then compare before/after on every PR.
Tessl:
- tessl-observability — measures where skills activate across real agent sessions, not where they are installed. The first capability here that measures agent behavior in production rather than in a test, and the distinction matters: a skill deployed everywhere and firing nowhere looks identical to a working one from any install-side view.
See Also
- artifact-perf-audit — the measured scorecard this pattern produces.
- stage-review — the stage where performance is assessed.