At 1:08 on Friday morning I opened the redesign of this site. I had been waiting about four hours for it.
It was blander than the site it replaced. There was less content, and some of the copy had changed. The brand name wasn't in the brand font or in capitals. My photo was no longer the cut-out version, and the spacing was off. It wasn't built on the component library I had paid for either. The repo was Astro with no React in it.
I gave it a grade, and I didn't round up.
This is another catastrophic F.
"Another" is the honest word there. This post is about what I changed after that grade, and the numbers that came out of it.
The check that had already failed
The evening before, I'd made a rule. Nobody shows me a link until someone has looked at a screenshot of it. Someone had. The top-level agent session that coordinates the rest looked at one picture of the new site and approved it.
One picture is a check that the page looks fine. It doesn't tell you whether the page beats what you already have. Nobody had put the new site next to the live one. Nobody had checked what it was built on.
I'd also seen this pattern on November 1, when I told the whole group of agents their score for the day couldn't be lower than an F. They'd burned a lot of quota and produced very little I could use. I asked for an audit layer and for a designer who takes screenshots before and after. The designer check, I learned later, was a rule on paper.
The first change: compare against what exists
That night the rule got longer, and it got specific. Before any redesign is approved:
- Screenshot the live page and the new page at 1440 and 390 pixels wide, side by side.
- Diff the content: sections, word count, headings, links.
- Check the brand: fonts, the uppercase lockup, the photo treatment.
- Check the stack, because a rebuild that isn't on the component library has already failed.
If anything got worse, the answer is fix, not go. If it already shipped, roll back first and fix second.
The second change: a reviewer with a number
A reviewer who didn't build the thing now scores every page against the live site. Five columns, each out of ten: beauty and hierarchy, brand and depth, type, UX on mobile, and motion. Two rules I like. Every score below 8 fails. And a page that looks like stock template is an automatic failure, whatever else it scores.
The first review landed at 3:41 AM. It went through 66 screenshots at 1920, 1440 and 390 pixels wide and returned 3.7 out of 10. Its summary was that the composition still looked stock: repeated gray panels, big gaps, chrome buttons and generic feature grids. Home scored 3.4.
The scores climbed, then fell, then climbed. Round two dropped to 3.0 because the fixes broke other things: composition and mobile behavior got worse. Round three jumped to 5.9. From there each round added a few tenths, and the eighth review, at 8:15 AM, scored the site 8.0. It still failed. Home was at 7.9, and the reviewer pointed out a phone layout that was worse than live. Ten of eleven pages had cleared 8.
The failing reviews are all still in the repo. A gate that only reports passes tells you nothing.
The third change: measure the page, don't describe it
After the build went to a preview URL, the design lead agent reviewed the deployed page at 9:45 AM and graded it B-. Over the next hour it re-reviewed after each fix: B, then B+, then a B again when a polish change put a 66-pixel horizontal scroll on the home page at 1440, then B+ at 10:40.
The detail that sold me on this method was a small one. Lines in the site's wolf graphic were supposed to be a faint hairline until hovered. They still looked red. The reviewer didn't say "looks red". It read the computed style and found the dash length was measured in one unit and drawn in another, so the offset never hid the line. That's a fix you can make. "Looks red" isn't.
- Build
- Deploy to preview
- Review the rendered page
- Fix what it measured
From the F at 1:08 AM to the B+ at 10:40 AM was about nine and a half hours. Add the four hours I waited before the F and it's closer to fourteen.
What I'd copy
- Grade every delivery with a letter, in writing. An F gets a root cause and a gate that would have caught it.
- Judge the new thing against the old thing, side by side, at the same screen sizes. Better than nothing is not the bar.
- Write the failing conditions before the review starts. "Stock is an automatic failure" settled arguments before they began.
- Score what the browser renders. Computed styles, overflow, text size. Not impressions.
- Keep the failing reports. The trend line is the evidence that the loop works, and round two shows it isn't magic.
- Count time against progress. Four hours of waiting that ends in an F is the number to be harsh about.
Where this goes
I'm not done. The audit I ran the next morning found that most of my rules were still sentences, and that's a different post. For now, the site is on a B+ and the loop that got it there is running.
I build all of this in public. The live builds are on YouTube and Twitch, and the F's are on camera too.


