Prove Nothing Changed
Give the agent an invariant and make it prove the invariant held. "Faster, and identical" is a task it can check.
The prompt
"Make this site's web fonts load faster without changing how anything looks." (Shu Ding)
The full prompt is a spec. It sets the invariant (rendering stays identical; any visual change is a regression unless approved), forbids side changes (leave font-display and fallbacks alone; no build step, a script run by hand with committed output), and orders the work: measure each font's size, glyphs, axes and features; find what the site actually uses, including inline styles and CSS side effects; subset to that; split each font into a preloaded core and a unicode-range extended file; then pixel-diff before and after on pages with code, italics, accents and CJK and explain every visible difference. It ends by asking for bytes before and after, the diff results, and every place it deviated from the request. Source: X/@shuding, 2026-10-29
Why it works
An optimization request without an invariant invites the agent to trade correctness for the metric. Stating what must not change, and how to check it, turns the task into one with a pass or fail. The agent measures first, so its later claims have a baseline. It explains each diff instead of hiding it. And the closing report makes deviations visible instead of buried in the diff.
The same structure carries over to audits. A responsiveness prompt forbids generic advice and requires running the app, reproducing each layout failure, tracing it to a component or CSS rule, making the smallest fix and retesting. Source: X/@shugarDadddy, 2026-07-14 A repo-audit prompt requires a file path and line for every claim, and an explicit "could not verify" when the agent cannot check something. Source: X/@meta_alchemist, 2026-06-09 @fredrikalindh's version is shorter: have agents test their own work with browser tools, and find another way to test, even listening for error sounds, when that is not enough. Source: X/@fredrikalindh, 2026-07-13
When to use it
- Any optimization where the output must look or behave the same: fonts, images, bundles, queries, refactors.
- Audits, where the risk is a list of plausible advice nobody checked.
Shu Ding's own warning applies to this whole category: read a prompt you found online before you paste it into your agent.
Which skill runs it
web-font-optimization is the font prompt as a skill: measure usage, subset into a preloaded core and a unicode-range file with a committed script, and prove it with pixel diffs. perf applies the measure-first rule to other speed work.
Once agents write their reports with screenshots of what they did, text-only reports feel like taking the agent's word for it. Source: X/@sawyerhood, 2026-07-16
Timeline
- 2026-11-03 | Page created from the web-font prompt and three audit prompts that demand proof. Source: X bookmarks, 2026-06-09 to 2026-10-29