← ClaudeAtlas

nebi-hillclimblisted

Read this before trying to make something measurably better, not after. Measure-first improvement loop for any engineering property with a number: page speed, load or response time, Lighthouse or Core Web Vitals, bundle or image size, frame rate, jank or layout shift, memory, build time, API latency, token or hosting cost. Use when asked to make something faster, lighter, smoother or cheaper; to optimize, speed up, tune performance, or cut cost; or when a numeric target is set, such as 'under one second' or 'Lighthouse 95'. This applies even when the request is a direct command to start optimizing — load it first, then measure. Not for fixing something broken or a sudden regression (nebi-debugging), an SEO audit or search ranking (seo-audit), conversion-rate or copy work (cro, copywriting), or refactoring without a measurable goal.
romanized/nebi-skill · ★ 1 · AI & Automation · score 74
Install: claude install-skill romanized/nebi-skill
# Hill-climb: measure, climb, lock it in The active harness's core §0 override always wins. This is the default method when you are deciding alone. Once there is a number that tracks what users feel, improving it becomes tractable. Most of the leverage is in finding the right things to measure, so measuring is step one of the climb, not a chore before it. ## 1. Pick what matters, then measure it - Start from the few journeys or outcomes users spend most of their time in (use analytics or field data when it exists), not from whatever is easiest to measure. - Make runs comparable: each measurement starts at a user action, ends when the result is usable, and separates client from server time where both exist. - Take a baseline before changing anything. Estimate each candidate change's gain in the target unit (ms, KB, euros), rank by estimated gain, and set the target from the sum. Leave room for opportunities you find on the way. ## 2. Make the number fast and trustworthy - Wall-clock and field numbers are what users feel, but they are noisy and slow to arrive. Iterate on a deterministic lab proxy: instruction or call counts, renders per interaction, layout and style recalculations, DOM mutations, bytes shipped, requests made, frames over budget under a fixed frame clock. - Prove a proxy before climbing it: improve it on one or two paths and show the real number follows. Drop a proxy that is flaky or does not track the user number rather than climb the wrong hill. - For we