The Record Goes Before the Review

On 23 September I asked for a targeted improvement of the Mac app's right-hand panel. What I wrote was:

The buttons at the top need to be spaced out better, and overall I feel like we need to think them out better … The only part which I feel is doing good right now is the Write Page tools? Everything else feels highly primitive and we need to improve it drastically?

The session did thirteen things, built them all, and ran out of time while its final review was still running. It ended without an account of what it had changed. I reviewed the changes on 14 October, three weeks later, from an account reconstructed from its transcript, and by then six of the thirteen reversed something I had decided.

01How it was done

Every pane of the right panel, on every page, was captured in light and dark first. AI agents then wrote down the Tools pane's design language, since that was the part I liked, and drew a design for each pane. Two competing designs were drawn for the row of icons at the top, and a judge chose between them. Separate agents built the changes in parallel, and the app built without new warnings.

Four reviewers were then sent in: one for taste, one for how it looked, one for regressions and one for speed. Each ran out of session after about 22 minutes. The final reports never came back, so I had no completed review to rely on. Some raw measurements survived, and I come to them below.

02What it had changed

Thirteen changes, each listed on 14 October with where it was made and what it reversed, if anything. Six reversed something I had decided. Among them:

• The icon row became one row with 8 points between everything and a thin rule only between groups.
• The assistant's words lost their bubbles and became serif prose on the column. People keep theirs.
• Messages always opens a conversation, the most recent. Conversations became an inbox with the last line, the time it moved, and an unread disc.
• Write's Chapter pane gained an editable Summary where an empty heading had stood. On 4 September I had said that a thinner rail is fine and nothing is invented to fill it. This reversed that.
• The Drafts rows went into serif, against the rule that a navigation row stays sans.
• Write got a lookup of the people and places the chapter names.

03My ruling

On 14 October I reviewed the changes in before-and-after pictures and decided to keep them. That decision did not settle the outstanding checks below.

I would like to say that I would have caught the six reversals sooner if I had been told. I am not sure that is true. I did not have a written account of those reversals when the session ended.

04The preview was not the app

The pictures first went to me from a preview tool outside the app. It substituted a symbol for the logo and a different font for the app's own, and I read the logo as having been removed. It had not been. The tool now supplies the real logo and the real font.

It is a small thing and I keep returning to it. A picture of the app is a claim about the app, and a claim made by a different program deserves the same suspicion as any other.

05What the later account left unresolved

• A probe on 23 September measured the selected inbox row's last line at 2.84 to 1 in Light, under the standard for body text. The measured sweep of 14 October then fixed the shared row: the selected row's last line went from 3.33 to 1 in Light to 8.9 to 1, with none under the floor.
• The Wiki's avatar rim had a workaround that was removed when a shared fix landed. I can find no record that anyone has since tested switching between Light and Dark with the app running, which is what it existed for.
• The speed reviewer's raw numbers had a Plot chapter switch at about 15 to 18 milliseconds before and 21 to 27 after. On 14 October I changed Plot so its Board and Table are redrawn only when something has changed, and measured the Board at 67 milliseconds before and 15 after.
• A line on the Desk said that a sample reader had suggested a cut in the harbour scene. There was no note behind it anywhere in the sample book, so it could open nothing. I said to give it one, and it has one now.

06The daily record

In July I started the daily record I mentioned in post two. My prompts are kept word for word, and each agent is supposed to record what it changed before finishing. This pass is the case it exists for. The transcript and saved files were enough to reconstruct the account on 14 October, three weeks after the session. That was when I reviewed the reversals.

I need a written account of each change before I can judge it. This session ended without one.

Thanks for reading.

J.H.

Next: Nineteen Hours of Not Fixing Anything →