Your AI Product Has a Craft Problem, Not a Model Problem

The models are good now. Good enough that the thing holding your product back is not the AI. It's the box you put the AI in, the words around it, and whether your team still has the taste to judge the output. Let me catch you up on what a handful of sharp design writers are saying this week, because they're all circling the same point from different angles.

The chat box is doing too much

Maxim Kich spent an hour walking his dogs and talking through an idea with an AI model. It followed him across design, engineering, physics, and back to an earlier hypothesis. Real thinking. Then he watched his best ideas disappear "somewhere between 10 tool calls, 3 deliverables, and 100 affirmations" that his question was fascinating. The intelligence is there. The container is a chatbot from 2015 on steroids.

That's the tension for anyone shipping an AI feature right now. Chat is easy to build and easy to ship. It's also a scroll of text where nothing has a home. Ideas don't accumulate, they slide off the top. If your product's whole interface is a message field, you've handed the user a smart tool and no place to keep what they make with it.

The model is only as good as your prompt

The fun part of these tools is watching them do work in seconds. Nick Babich ran the same product research in two Claude setups and found AI-assisted research beat human-only work on both quality and speed. But he starts every project by answering three plain questions before touching a prompt: what do you want to know, why, and how will you use it.

Same story with data visualization. Theresa-Marie Rhyne used Perplexity to build diverging color schemes, and the good results came from examining and modifying the specific hex codes the model suggested, not from taking the first answer. The AI proposes. You still have to know a bad midpoint color when you see one. Speed is real, but it only pays off if someone on your team can direct and correct the output.

Craft is the new hiring bar

Here's where it gets uncomfortable for team leaders. Tom Scott pulled from an analysis of 140-plus hiring calls, and the top reason designers get rejected at AI-native companies is craft. The fear is real and specific: if everyone delegates execution to AI, everything starts to look the same. The people who stand out have the taste to judge and steer the output, not just generate it.

The role itself is shifting under your team. Technical fluency is now assumed, not a bonus. Handoffs are functional code, not static Figma files. One company Scott spoke to is hiring only Design Engineers, no Product Designers, because they think the old PM-design-engineering triad is breaking down. And one hiring manager said the line that stuck: "When anyone can build a product with AI, brand is what differentiates you." That's the part AI can't fake yet.

The words carry the product

While everyone chases the models, the oldest lever is still sitting right there. Nick DiLallo, who has written for billions of users, opens with a plain fact: most interfaces are mostly words. Take away the text and screens stop making sense. Fixing the writing can drop support tickets by the thousands and lift conversion.

His warning lands harder in an AI product. Making a thing "conversational" can just make it annoying to use. A button that says "Pay" beats one that says "Ka-chingggg." And writing quality signals product quality: you can't ask someone to trust you with their private data if they can't trust you with your commas. When your product talks back to users all day, every sentence is a promise about whether the whole thing works.

The deep cut

DiLallo's math is the thing to carry into your next review. He points out that a ten-second fix, multiplied across a billion sessions, is measured in years, not minutes. That reframes every "it's just wording" argument your team keeps having. At AI scale, a bad label or a clunky affirmation isn't a small miss, it's a tax you pay on every interaction.

So the move on Monday isn't a new model or a new chat feature. It's putting your sharpest editor and your most opinionated designer on the smallest surfaces: the confirmation text, the error message, the default color, the button verb. That's where craft compounds, and it's the one place AI can't yet out-taste your team.

Three questions for your team

1. Where in our product does the chat box carry more than it should, and what should have a permanent home on screen instead of scrolling away?
2. Can everyone on our team judge and correct AI output, or are we shipping first drafts because nobody has the taste to push back?
3. If anyone can now build our features with AI, what makes ours feel like ours, and who owns that brand and voice across every screen?