Vibeleaderboard
← All Intel
Intel / video

How an Anthropic Designer Prototypes Real Features With Claude

Source
youtube.com
Author
Peter Yang
Date

Context

Nate Parrott, a designer at Anthropic whom Peter Yang introduces as the designer behind Claude Design and Claude Motion, describes a way of prototyping that starts from the product's real code instead of a separate design file. In the interview, he says he has Claude read the codebase and redraw the current screens on the Claude Design canvas, the space where Claude lays out designs, slides and prototypes. He then asks for many variations, narrows them with short instructions, and when a direction is settled, tells the same Claude Code session to build it. Because the session already holds the codebase and the design, he says, very little is lost between the mockup and the implementation. Claude Design can now be used from inside Claude Code; Parrott says it began as a separate internal tool he built for himself.

His example is an early idea for Anthropic's personal , which the automatic captions name Duet. The team wanted Duet to suggest many tasks it could take on, since a user might need to see about 20 to find one they wanted. Parrott typed a rough note on his phone while waiting for the subway, pasted it into a remote Claude Code session attached to the real codebase on the walk home, and asked Claude to recreate the current screen and propose options. The next morning the canvas held the existing daily briefing next to several layouts. He refined them mostly by typing changes such as a denser list with no dividers, kept every version on the canvas for comparison, and handed screenshots to the team. That design was not built, but he says Duet as a whole was made by putting ideas into an internal build, watching whether people used them, and replacing what nobody used.

Much of his advice is about getting Claude past its default look. "You get out what you put in," he says: he supplies screenshots as a mood board, examples of pages he likes and his own past work, and for a new design asks for about ten options with different fonts, colors and styles, sometimes starting with plain wireframes so layout is settled before visual detail. Left alone, he says, the model tends to land on serif text, an off-white background, orange accents and bars along card edges, believing each time that the style is original. Anthropic's prompt for Claude Design therefore names the fonts and colors the model over-uses and tells it to avoid them unless asked. He adds that design quality is hard to test automatically, because a model without good taste cannot grade taste, so the team checks basics such as overlapping text and repeated fonts and otherwise looks at the outputs.

The change Parrott describes is in the speed of building, not in learning whether a feature is wanted. At The Browser Company, which made the Arc browser, he says each idea took a couple of weeks to prototype and the team spent about a year in pre-production; at Anthropic a comparable round takes one to two days. Putting a prototype in front of people and seeing whether they use it, he says, has not sped up. That matches the prototype-first, internal-adoption process another Anthropic employee described for the Claude Code team. This account comes from automatic captions of the video, which garble some names, so product names here follow the captions. The on-screen demos were not seen, and this is an Anthropic designer describing his company's own products. Yang also posted the Arc comparison on X.

Key takeaways · AI-distilled
  • Anthropic designer Nate Parrot avoids look-alike output by asking Claude for many options at once, around 10 with different fonts, colors and styles, then asking for new variations built from the parts he liked.
  • For existing products, he has Claude read the codebase and recreate the current UI on the canvas before designing, instead of maintaining long-lived design files. He then builds in the same Claude Code session, so little is lost between design and code.
  • Parrot says the Claude Design prompt tells the model which fonts and colors it tends to collapse to and to avoid them unless asked, because the model cannot see its own biases.
  • He rates Opus 5.5's motion graphics as okay rather than great: animations often change too many things at once, so he prompts for reading time and a single point of focus, and uses timestamped canvas comments for fixes.
  • To tune a , he spawned fresh Claude Code sessions given only the skill and a prompt to make a reel for a boring feature, then judged the outputs by eye. He prefers yes/no checks to 1-5 rubrics and says design quality is hard to evaluate automatically.
Terms in this piece · Glossary
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
  • eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
  • agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
Read the source youtube.com
More from Peter Yang
Recommended reads
Comments

Checking sign-in…

Loading comments…