I rebuilt more than fifty screens in Figma in a single day. By hand, with other tasks running in parallel, it would have taken a month.
The project was this: the design system existed, but production and design had drifted far apart long ago. Production had to be captured whole and the screens rebuilt on the existing system.
It didn't work that way before. The agent made things up, ignored what was already there, and invented rules that don't exist as it went.
The most common example is fonts. The agent will burn a million tokens trying to apply a font that is right there in the system, and finish by redoing the whole design in Inter. It sees the methods in the API. It pokes at them. But it doesn't know how to work with them.
Wording the request doesn't fix this. Plenty of people still believe that a well-written prompt gives a good result, maybe even good design. No. What fixes it is knowing the case and checking the result.
The knowledge piled up on its own. Agents have a built-in retrospective: the task ends, you go through what worked and what didn't, and it gets written down. Some of it I said directly: you see it grinding on the same thing, or repeating a mistake a second time, and you tell it to remember. That came to 378 entries. It surprised me — I didn't think there were that many.
Here is what one looks like. Not "be careful with components", but: every element gets exactly three permitted actions — place an instance, update the design system, make a new component. Your own frame that looks about right is forbidden. If none of the three fits, stop and ask the human.
That day grew out of this. I wasn't the one assembling the screens.
But everything still had to be reviewed. There were inaccuracies, and there always will be. Only I wasn't the one fixing them either: I made the agent fix its own mistakes.
So I put it in open source. When you publish something, you start working on it more: on your own you can slack off somewhere, let something slide. I thought about how to make money on it and decided not to sell it. Going further with it interests me more.
Inside are five skills for Claude Code. One plans screens and states before anything is drawn. The second reads the file and reports what is in the design system. The third designs and edits screens. The fourth is those entries; it loads before any write to your file. The fifth repairs detached variables.
Get it: github.com/schudarin/chudarin-figma-skills. MIT, install instructions in the README.
I need feedback: what broke for you and what the rules didn't cover. And if you write skills like these, write to me. Something might come of it for all of us.