← Writing

Left alone, our AI agent reinvented the wheel. So I made it look first.

Twenty prototypes in twenty minutes were the easy part. Making it study what already works, and why, took two days and three rules.

Building · Drafting · Nov 2026 · 6 min read

By

TL;DRAn agent asked for a redesign will happily reshuffle the same parts. People find familiar patterns easier, so I made mine research what already works, and why, before drawing anything. Pair that with a design system that has hard rules and it goes from an issue to merged code in an afternoon. Real users still decide how much of it you need.

At 10:52 on a Tuesday in October, I posted in our frontend channel that the admin screens wasted space. The calendar spent 287 pixels on a title, a month picker and a legend before it showed a single day.

Our team runs an internal Slack agent. It’s Claude-based, and it can read our repos, run a headless browser and publish internal pages. I could have asked it for a redesign and gotten one in minutes, and that was the problem. Ask an agent for a better layout and it moves the same parts around until something looks fresh.

Fresh isn’t what you want from an admin tool. Jakob Nielsen wrote it down in 2000: users spend most of their time on other sites, so they expect yours to work the way those do. Familiar patterns are easier to use. So the job was to find out which calendars already work, and why, and build ours on that.

Getting the agent to work that way took two days and three rules.

Every variant had to name what it copies

My message asked for three things: variants instead of one answer, references from comparable products instead of taste, and priority. Is the main thing on the page actually the main thing?

The agent captured the real app first and measured the problem: the calendar grid started 287px from the top, and the lesson builder’s content at about 640px.

287px of headerBefore
104pxEvery variant
Where the calendar starts on a 900px screen. The rule from the agent's principles file: everything above the main module fits in 104px.

Then it wrote a foundation before any variant: our existing tokens, a shell copying the real sidebar, mock data shaped like the real types, and a short principles file. One primary module per page. A 56px header and at most one 48px toolbar above it, so 104px in total. Fixed sizes only. One filled button. Colour only carries meaning.

Only then did it fan out parallel sub-agents. Each variant had to name a real product, a concrete layout and the interactions that had to work. The calendar came back as Google Calendar, Notion Calendar with a side rail, and a timeline like Float. Twenty clickable variants across seven pages, each next to a screenshot of the page as it is.

The foundation is why twenty prototypes from parallel agents looked like one design system and not twenty opinions. And 287 → 104 meant nobody had to argue about whether it “felt cleaner.”

Red-teaming against the user stories

Looking good is the easy part. I asked for a second pass against the use cases and user stories.

It pulled the five personas from the product’s own planning docs, walked each variant through the end-to-end flow, and reran everything at 1280×800, the laptop someone actually carries to an on-site training. It found ten gaps that no amount of visual polish would have fixed, among them:

  • No calendar showed which trainings still needed a lesson plan.
  • The QR code for the day was behind a hover-only icon. That fails on a touch screen at the venue.
  • Billing happens per training and per month, but the tables were per organisation, so you had to open each one to find the work.
  • The shared sidebar had dropped logout, theme and font size. Nobody notices those until they’re gone.

It built a fourth variant for the pages that needed one, mostly by combining the best parts of the others, then gave one recommendation per page and a build order. The whole thing, both passes, took about 35 minutes and $22.

That evening I asked it to keep the approach as a skill: a markdown file with a name, a trigger and a procedure, which loads into any session that matches.

The day it skipped the research

The next day I used the skill on a list page with too many status colours. Three minutes later I had a pull request against the real frontend repo, with every status turned grey. Nobody had looked at alternatives yet.

I didn’t want a PR yet. Close that. Show me your usual redesigned prototype designs.

The new variants were fine, but the references felt familiar in the wrong way. They were the products an AI model would name first. So I asked:

Did you really run my redesign skill?

It answered “only partly.” It had picked the references from memory and skipped the red team.

Then it did the research properly, and this is the step I care about most. It wrote the page’s information architecture as one line, “a list of things you made, plus a visibility setting, a review status, and favorite and ⋯ menus,” and searched for products with that IA, not products in the AI space. GitHub’s repository list, where visibility is a grey “Public” badge and the only colour is the language dot. Hugging Face Spaces. App Store Connect, where yellow means a process is running and red means you need to act. YouTube Studio. And Shopify Polaris as the counter-example, because it does use coloured badges.

That research turned up a real bug in our product. One state the owner needed to see wasn’t shown to them at all, and every comparable product shows it. Reshuffling our own layout would never have found that. Reading how other products handle the same states did.

Three rules it can’t skip

I asked it to harden the skill. It now opens with three rules, each written because it was missed once:

  1. A redesign request means prototypes, not a PR. No repo edits until someone picks a variant.
  2. Every step runs, even for one component. Scale the counts down, never the steps. If a step is skipped, the first line of the reply says so.
  3. References come from research, not memory. Describe the IA in one line, find six to eight products with the same IA, including ones outside your domain, and include one counter-example. Every variant points back at that list.

Each rule carries its date and the reason, so nobody “simplifies” it away later.

The third rule is Jakob’s Law, made into a step an agent can’t skip. It’s also how we’ve built most of what we ship: a learning platform that works like Duolingo, a store that works like the App Store, a studio that works like ChatGPT, vibe coding that works like Lovable, and admin screens that work like other cloud platforms’ dashboards.

From an issue to merged code in one afternoon

The research decides what a page should be. The design system decides what it’s made of.

Ours has components, plus rules an agent can check. The theme fixes the sizes: a 56px header, 32px controls, a 220px sidebar, a type scale of 12, 13, 14, 16 and 20, and corner radii of 6 and 8. The admin has its own short skill that says what was asked for and how to check it, and a decision log that links each product decision to the Slack message behind it.

When the redesign was picked, there was little left to decide. A teammate built the chosen variants with the agent that same afternoon. Seven pull requests, the first opened at 13:43, all merged by 18:25. Plus 11.4k lines, minus 3.0k. The prototypes used the same tokens as the code, so most of the build was translation.

One gap: the spacing steps and the 44px row height were deferred in the first pull request and never made it into the theme.

What survived two days of real users

Then people started using it, and about half of it changed within 48 hours.

Survived
  • The fixed sizes: header, controls, type
  • The calendar's structure
Reworked
  • The dashboard, reversed the same day
  • The organisation drawer, now one table
  • Accounts, cut back to a single table of people
  • Organisation detail, cut down again two days later
The training admin, 48 hours after the redesign merged.

Three of the four reworks cut things back. The foundation held; what changed was how much sat on top of it.

The red team had five personas, but personas don’t push back. The research told us how each page should work. Only the people using it could tell us how much of it they needed.

[Sean: what you’d do differently next time, now that you know half of it got cut back.]

Use it

The skill is public, rewritten without anything specific to our tools. It works with Claude Code, Codex, Cursor and the other agents the skills CLI supports:

npx skills add spark33/spark33.github.io --skill ui-redesign-variants

The source, with the kit (tokens, shell, the capture and probe scripts, the index builder), is on GitHub.

Drafted with my agent from my own draft and an interview; edited by me.

Disagree, or tried something like it? Write to me. I read everything.

RSS

Keep reading