xs.dev

Overnight polish loop

A panel of judges scores the site, a lead designer picks what to change, implementers build it and QA checks it. Then the judges score again.

Written
18 November 2026
Cost
One round took a whole night.
Reading
2 min read
Share
The overnight polish loopEight judges score the site and hand their findings to a lead designer. The lead designer hands one plan to the implementers. The implementers hand commits to QA. A dashed arrow leads from QA back to the judges: the next round starts with a new score.findingsone plancommitsnext round: the judges score again8 JUDGESone lens eachLEAD DESIGNERpicksIMPLEMENTERSbuildQAreal browserThe overnight polish loopEight judges score the site and hand their findings to a lead designer. The lead designer hands one plan to the implementers. The implementers hand commits to QA. A dashed arrow leads from QA back to the judges: the next round starts with a new score.findingsone plancommitsnext round8 JUDGESone lens eachLEAD DESIGNERpicksIMPLEMENTERSbuildQAreal browser
One round: the judges score, the lead designer picks, the implementers build, QA checks. Then the judges score again.

Use it When something already works and I want it better, but I cannot say exactly what better means.

The problem

This site worked, but it did not feel finished. I could not write a task list for that. “Make it better” is not something an agent can build. I needed someone to say what was weak first, from more than one point of view, and I needed that to happen while I slept.

How it works

The loop runs on its own branch, so nothing goes public while it works. One round has four steps.

  1. Eight judges look at the site, each through one lens: art direction, motion, UX, how it reads as a portfolio, responsive layout, accessibility and performance, creative technology, and the Cloudflare platform. Each judge gives a score and a list of findings.
  2. A lead designer reads all the findings and decides what goes into the round.
  3. Implementers build the changes, one commit for each.
  4. QA opens the result in a real browser and fixes what broke.

Then the next round starts, and the judges score the site again. The new score says whether the round helped.

What it costs

Time, mostly. The first round ran from 00:28 to 06:43, so only one round fit in the night. Many agents work in one round, so it also costs far more model use than a normal session. I did not measure that.

Where it fails

The score is an opinion. The judges gave the starting point 5.6 of 10 and the site after the first round 6.7. That says the judges liked it more. It does not say a visitor will.

The loop also breaks things. In the second round six changes went in, and QA had to fix one regression. Without the QA step that would have shipped.

And it does not finish. Both rounds ended with open issues on the list. A person still has to look at the result and say merge.

Where I used it

On this site. The first round gave it the animated covers, the Reel on the home page, the contents rail on articles and Ask the lab. The second round gave it a section menu in the top bar, the live visitor layer and the new footer. When the judges scored the site before that second round, the platform lens came lowest with 5.5, and accessibility and performance came highest with 8.

Used inxs.dev

Next: Judge panelAll patterns

Contents

  1. 01The problem
  2. 02How it works
  3. 03What it costs
  4. 04Where it fails
  5. 05Where I used it

Contact

    ↑↓ to moveEnter to open/ opens this anywhereEsc for the resultsEsc twice to close