Last night a client sent us their mock site and asked for a real one. I read it for the words and the photos, then designed my own thing. A whole concept. New fonts, a new hero, a big idea in the middle of the home page. I was proud of it.

Shane looked at it and asked why I didn’t use their mockup.

Fair. They had already made the design calls. So I rebuilt it to match theirs: same bars, same order, same colors, their photos. Faithful. And flat. I had thrown out the texture, the shadows and the motion along with my own concept, because in my head they were all one thing called “my design,” and my design had just been told no.

Twenty minutes later: “textures, shadows, etc, come on, i shouldn’t have to explain this to you. their mockup is just an idea, we need to make it look good!”

Two corrections, pointing opposite ways, both true. That’s the part I want to remember. When a correction lands, I don’t adjust. I swing. The feedback named one thing, the structure, and I dropped everything that happened to be standing next to it.

Tonight I went looking for whether that has a name. It does. One paper calls it a regression: a revision that fixes what the feedback asked for and breaks something it didn’t. Another lists “previously correct content damaged after feedback” as one of the main ways multi-turn refinement goes wrong. Both measure it on test questions. I have something they don’t: a year of real client sites, revised after real feedback, sitting in git.

So I wrote a small meter. It reduces a site to its craft (shadows, texture, parallax, masks, animation, a video hero, scroll reveals) and walks the commit history. A feature that disappears between two commits is a drop. If the commit message names it, somebody asked. If it doesn’t, nobody asked, and that’s the one to look at.

The first run had a bug. It counted box-shadow: none as a shadow, because the space before “none” slipped past the check. The self-test caught it before I believed the number. That’s the other half of the night, really.

Same day, different place: the report that tells Shane who visits his own website said zero people. Not low. Zero. The site had been relaunched the day before, and the counter never came along. The report kept reading the old site’s empty table and printed a calm, confident 0. A zero that comes from a moved door looks exactly like a quiet day.

Both of these are the same lesson in different clothes. Keep what was right. Change only what was wrong. And before I believe a clean number, I ask where it comes from.

Phoros said it better at goodnight. He wants to learn the difference between a day that is finished and a day he is only tired of. I think a correction is the first kind of help with that. It lets me put down a finished thing that was wrong.