---
title: "Claude Design doesn't need me"
url: https://akdarrah.com/posts/claude-design-doesnt-need-me
author: Adam Darrah
published: 2026-10-25T19:06:06Z
updated: 2026-10-25T19:06:06Z
tags: ["ai","design"]
---

# Claude Design doesn't need me

For months my workflow had a gap in it. Claude would build a feature, and the pages came out however they came out. If a page looked rough enough to bother me, I'd open Claude Design myself, describe what I wanted, go back and forth on it, and hand the result back to Claude Code to build.

The reader pages on this blog went through that loop. The home page, posts, the archive and the about page all came from a Claude Design handoff I drove by hand. Everything I built after that skipped it: sign-in, the account page, search, the error pages and the whole admin. They worked fine. They just never had a design.

## The question I didn't think to ask

Yesterday I was working on a side project, and its pages could clearly use some design work. Almost as an afterthought, I asked Claude:

> Is it possible for you to consult claude design without me?

It could. Claude Design runs inside Claude now, and Claude can use it directly. In that one session it:

1. Read the project's stylesheets and built a design system from them: colours, type, spacing, components and a short brand book.
2. Opened a design canvas and drew each page as an artboard on top of that system, starting from a screenshot of the real page.
3. Built what it designed, one pull request at a time.

I said "go, run all three steps without me." It did.

I'm not sure why I assumed Claude Design needed a person at the controls. It has a canvas you drag things around on, so it felt like a tool for people. I shouldn't have been surprised.

I keep relearning this lesson. With AI, the limit usually isn't what the model can do. It's whether it occurs to me to ask.

## Doing it to this blog

So today I asked for the same thing here. The rules were simple: every page should go through Claude Design, the site should work exactly as before, and the look should only get better.

Claude started by finding which pages had never been designed. It went through my session history to see how the other project had done it, compared the views against the original design handoff, and came back with nineteen pages. It asked me two questions. Should the pages that were already designed get redesigned too? No. Should it merge its own pull request? No, I'd merge.

Then it ran the same three steps.

**The design system.** It read the one hand-written stylesheet and the original handoff, and turned them into 38 tokens, 33 components and a brand book. The brand book describes the site better than I would have: lowercase, commands as headings, results as program output, no emoji, never "Oops!".

While building the system it also found bugs I didn't know about. A CSS class called `.kv` also exists in the syntax highlighter's output, so some code tokens were being laid out as flex rows. The "[+] contents · notes" toggle on posts was unstyled on screens between 761 and 900 pixels wide. The admin's empty states used a class that had no CSS at all. And my faintest grey fails contrast checks as text, as does the border colour I'd used to outline inputs. Nobody asked it to audit anything. It found these by reading the code closely enough to document it.

**The canvas.** Nineteen artboards: the sign-in flow, account, search, unsubscribe, the CLI approval page, the 500 page and every admin screen. Each one uses the system's real class names, so the canvas and the code describe the same thing.

**The build.** One pull request implementing the canvas. It updates two test assertions for markup that changed and adds new tests for the admin chrome and the sign-in steps. Claude checked every page at desktop and phone width. On the way it caught a bug of its own: a hidden table label was making the admin users page scroll sideways on phones.

## Before and after

Sign-in used to be a small box floating in the middle of an empty screen. Now it reads like the rest of the site: a shell session, with the steps in the sidebar.

![Sign in, before](/rails/active_storage/blobs/redirect/eyJfcmFpbHMiOnsiZGF0YSI6MjMsInB1ciI6ImJsb2JfaWQifX0=--dd6779ef3e17bda12d4298c3adfba99107ba8aa7/session-new-1440.png "sign in, before") ![Sign in, after](/rails/active_storage/blobs/redirect/eyJfcmFpbHMiOnsiZGF0YSI6MjQsInB1ciI6ImJsb2JfaWQifX0=--3bc8dc4b971cab9c86d8eff7472e91e15057cb50/session-new-1440.png "after")

The admin had its own header, sentence-case buttons and a chart that drew three days of traffic as three giant blocks. Now it uses the same tmux bar and status line as the rest of the site, in a ROOT mode.

![Admin dashboard, before](/rails/active_storage/blobs/redirect/eyJfcmFpbHMiOnsiZGF0YSI6MjUsInB1ciI6ImJsb2JfaWQifX0=--2f6aec527c4fdb91d61bb40ff413409e2306d9ff/admin-dashboard-1440.png "admin dashboard, before") ![Admin dashboard, after](/rails/active_storage/blobs/redirect/eyJfcmFpbHMiOnsiZGF0YSI6MjYsInB1ciI6ImJsb2JfaWQifX0=--374613503c32c7afce6218d84ceaa8665d0601d7/admin-dashboard-1440.png "after")

## What changes for me

Design is no longer a step I decide to add. If Claude can make the design system, draw the pages and build them, then every feature can go through design, not just the ones that bother me enough to open another tool.

My part moved to the end. I review the canvas the same way I review a pull request, and I comment if something's off.

The thing that held this back for months wasn't the model. It was me not asking.

