I art direct my agents with comments on a canvas
I stopped describing visual problems to agents in chat. I leave a comment on the canvas, on the screen where the problem is. My agents draw the screens on Paper, a design canvas they can read and write through an MCP server. A fresh agent reads my comments, fixes the design and resolves them. Only then does anything go to code.
The design lives on Paper first
In August I was building a tool for landlords, and the pages got more complex every time the agent added something. I was overwhelmed by them. So I moved the whole interface onto Paper, to judge the UX and clarity there, change it there and fix the code afterwards.
On the canvas it was a wall of text. We want to take stuff off people’s shoulders, and we were bombarding them with text, text, text. Where it could be a calendar, it should be a calendar. Where it could be tiles, tiles. Where it could be icons, icons. I had the agent put captions and notes on the screens saying what’s where, so I could check everything at a glance.
”On paper” means Paper
One agent heard “on paper”, built HTML and checked it in an agent browser. I had to spell it out:
on paper means paper mcp, when we doing the design we use paper mcp, thats whats on paper means
The build order is fixed now. Tokens and a design system come from the landing page first, one Opus agent checks the spec, then another builds the interface on Paper. Impeccable, a design skill, is there to separate real attention to detail from slop. Hair lines and the on-track dot are AI tells, so they’re out.
The comments are the brief
I don’t translate a visual problem into a sentence anymore. A new agent with a clear context reads the comments, uses Impeccable and works from them. It needs to pay attention to details and do good work.
Comments queue up. If an agent is still running, I leave them anyway and it picks them up when it finishes.
Choices work the same way. When a menu needed a decision I didn’t want variants in code. I wanted them drawn on Paper, each beside the original, so I could make the call from there.
Nothing gets ported until I approve it
That’s the rule, and agents keep testing it. A menu got ported to code before I’d looked at the new version on Paper, the day after I’d said “wdym port? first i need to see how it changed on the paper then confirm then port.” It happened again:
why does it port if i didnt accept the menu on paper yet? thats the rule if we do new stuff i first need to confirm it on paper
Copying an app to Paper 1:1
For my media canvas app I want the same loop built into the product. The spec says the point is collaboration with the agent: you can comment, the agent checks the comments, adjusts and resolves them, and you can see whether the agent is working right now and where.
To start I wanted the app’s existing screens on Paper, 1:1, so I could comment on them there, finalize the design there and then update the app. The first agent over-read “1:1”. It took screenshots and moved real content over, when 1:1 meant only the design, which it could mostly take from the code. I stopped it and deleted what it had made.
Grok in Cursor was faster and did a good job. It copied the app well and I left my comments straight away. An Opus agent with Impeccable works through them with one brief: as little cognitive load as possible.
The file gets huge
The Paper file grew until I couldn’t tell what was going on in it. An agent now reads through everything, inventories it, marks what’s stale and archives most of it. The one rule for the cleanup is to delete nothing: resolve comments and move things into the archive.
Background agents get stuck too. Some post a final message and then do nothing, or sit on a blocker. When one starts doing stupid stuff, I have it clean up and start another fresh-context agent instead of arguing with it.
The media canvas app’s design is on Paper now, with my comments on it.