Claude Design rocks and Claude Design sucks. Let's talk about it. Last week, I m
By Allie K. Miller · April 28, 2026 · Curated by George's Blog
Claude Design rocks and Claude Design sucks.
Let's talk about it.
Last week, I made an entire game in Claude Design inspired by the Longyou Caves in China (why is there no soot!), and we all gathered around the TV, Airplayed it, and shouted at the screen to solve ciphers.
Was it fun? YES!
Was it perfect? No!
Claude is still rough. Here are weak spots where I'd go to another product, work extra hard, or hire a human:
1. Anything SVG from scratch (ex: caves or bat characters) - the illustrations Claude makes are comically bad. The "bat with a flashlight" looked like a fat clown fish stuck in a purple hat wearing a misshaped pearl necklace
2. Accessibility decisions, like font sizes or color schemes - if you prompt Claude to fix these, it can look up best practices and update the design decently well. But as a colorblind person with strong contacts, I would trust a human over AI any day. One example it kept failing on was what color to highlight text with
3. Design principles - strange font selection and font family mismatch (how many serif fonts can one gal handle?) + the game toolbar overlapped and hid about 1/3 of the screen
4. Human-centered basic design - there were so many obvious features Claude missed. Ex: the cigarette in one screenshot below was just a dot and when we had to 'log' the find, we logged it as a cave drawing. Claude came back and said "false! it's a cigarette!" but gave zero design or caption signaling. Very obvious gap if you're a human user of it, not obvious if you're an AI designing it. The user still owns UI QA
5. Human-centered delightful design - it'll do what you ask. Will it add extra levels, a fun mascot, engaging buttons, or whimsical captions to drive the game forward? Eh. The game was timed and we asked it to add a 'health' meter, which it called 'stamina', but because it set it up as a timer too (not health points), the game had two timers running
6. Logic - we had to label cave objects as one of four categories - authentic, anachronism, forgery, or 'natural/modern'. Modern is also anachronistic so that made no sense. We labeled one object as 'natural' and the system said "nope that's actually a ghost!" and marked our answer wrong...but ghost wasn't one of the options
7. Iteration and tradeoffs - I want AI systems to consider tradeoffs. If we're dealing with a massive design overhaul, I'll answer 10 questions. If we're changing one font, 0-2. We asked for a kid-friendly version; Claude asked 0 questions and built for a user too young. It lacked reasoning/empathy and didn't make the right tradeoffs
8. Token usage - mine was fine but I saw several posts saying it killed theirs, so worth calling out and checking with '/usage' in Claude Code
It didn't replace a PRD process or the deep iteration I do on product building but it made me slightly less likely to recommend non-engineers look into Lovable as a design prototyping buddy.
'Claude the Everything Machine' is getting more vertically integrated every day.