Wink Pings

Claude Thinks It Invented Every Rounded Corner, So an Anthropic Designer First Has It Draw Wireframes

In Peter Yang’s show, Nate Parrott demonstrates how to use Claude for design without falling into AI slop. Draw a wireframe first, ask questions first, and only then let the model touch fonts and colors.

Anthropic designer Nate Parrott demonstrated on Peter Yang’s show how to get Claude to do design without falling into AI slop. The key is not just swapping prompts, but first pushing the model back to the problem itself.

Peter Yang quoted one of Nate’s lines on X: a good designer asks what problem we’re solving and which interfaces people are already familiar with, and you can have Claude ask that too. Nate’s approach is more concrete: sometimes draw a wireframe first, so Claude doesn’t plunge straight into fonts and colors. For example, when building an inbox, first have it think about other inboxes people have used, find one with a suitable structure, and then draw a wireframe using the same structure.

Behind this is a counterintuitive judgment. Peter Yang quoted Nate in his tweet saying that Claude doesn’t know what its own design biases are; in Claude’s mind, it has only ever done this once, so it thinks it is completely original. A commenter added a sarcastic reply: every Claude design thinks it is the first to discover rounded corners plus purple gradients. Another person said a wireframe is a set of reins for a model that has just discovered fonts. Blunt, but it makes the problem clear. What the model really lacks is constraints and references; aesthetic vocabulary doesn’t help.

Even more counterintuitive is friction. Traditional design says don’t add friction, but Nate said on the show that the discovery of Claude Design is that you can ask a lot of questions, and users actually like it. Some commenters were surprised; they had assumed users could tolerate questions at first but would get annoyed once work got underway. Another commenter who works on agents said most agents choose to guess when they should ask. This observation applies beyond design. Someone mentioned legal memos: AI drafts always default to one structure, and asking it for two different outlines first can break that. Another person used an odd trick for writing: telling the agent the draft was written by a cousin suddenly let it see that all the drafts had the same flavor.

On the show, Nate demonstrated several things: designing a Duet feature from phone notes, turning a spreadsheet into a live dashboard, and then turning the dashboard into a Claude Motion video. The show’s chapters also list four traits that make design look AI-made, and how to test design skill when AI cannot judge taste. These details are more useful than a line about improving aesthetics.

Peter Yang’s full show is here: https://www.youtube.com/watch?v=604T-74s3KE

In the main clip, Nate explains the process of constraining Claude with wireframes and known interface structures:

Full show:

Peter Yang’s original post: https://x.com/petergyang/status/2122015592001708110

A takeaway ordinary users can also apply: before asking AI to make an interface, a page, or an ad image, first give it a layout users already understand, then give it a specific confusing scenario. A commenter gave an example: “I can’t find how to cancel” in a support ticket is a good prompt; it gives Claude a moment when a user is stuck, before fonts and colors enter the conversation. Draw the wireframe first, then talk about beautification. Most rework happens in structure; color rarely takes the blame.

发布时间: 2026-11-16 10:28