Steve Jobs once said the real problem with Microsoft wasn’t the engineering — it was that they had no taste. I used to read that as a jab. The more I use AI for the actual work I care about — research, writing, small tools I build for myself — the more I think it was the whole point.
Taste isn’t a vibe. It isn’t “design sense” in the narrow, aesthetic meaning. Taste is the compressed output of everything a person knows — their experience, their empathy for the people they build for, their feel for what actually works and what doesn’t. It shows up in a thousand small decisions: what to include, what to leave out, where to push, where to stop. None of it is copyable.
And it’s decisive.
A concrete example. Recently I spent about half an hour putting together a briefing on REITs in China — public REITs, private REITs, “quasi-REITs.” Simple-sounding topic, but almost everyone who talks about them confuses the three. The differences are subtle, and subtle is where every misunderstanding lives.
The report was good enough that people kept asking me what AI I used. That wasn’t the interesting part.
What I did: force the AI to quote only the direct wording of government policy and regulatory text — no second-hand summaries. Draft the distinctions from those primary sources. Then ask the AI to compare the categories side by side, because I already knew that was where the confusion sat. Then run the whole thing through a small app I built for myself — FormalDoc.app — which turns Markdown into the government-document format Chinese readers expect when they open a policy brief.
Primary sources only. Compare, don’t describe. Format like the reader. None of those standards came from the AI. They came from years of watching where readers trip up. That compression was mine. The AI was the tool.
The REIT story isn’t special. It’s a visible version of something AI has made obvious across every kind of work: the text is doing all the work. Not the model, not the technique. What you ask for, what you’ll accept, what you push back on — those decisions are where every output is actually made.
Everyone has the same tools. The same models. The same prompt examples floating around. But watch what people actually produce — the gap is enormous. One person asks an AI to rewrite a paragraph and gets something sharp, specific, alive. Another asks for the “same” thing and gets mush.
The difference isn’t prompt-engineering technique. It’s what you ask for and what you’re willing to accept.
When you have taste, you can tell within two seconds whether an output is right. You feel the gap, push back, rewrite. You converge. When you don’t, you can’t tell. You accept whatever comes out, or you wander through variations that all look “kind of fine.” The tool is identical. The ceiling is not.
This is why AI makes taste more valuable, not less. The floor of competent output has risen. The ceiling of excellent output has risen faster. And the gap between the two is almost entirely a gap in standards.
I don’t think the most sophisticated systems in the world are shaped by their engineering. They’re shaped by the lead designer’s taste — refracted through a thousand small judgments an outsider would never notice, but that the user feels the first time they touch the thing. That’s the part nobody can copy by reading the source code.
So if you’re worried about AI flattening the field, I wouldn’t be. It’s doing the opposite. It’s taking the hidden thing that used to only show up in the final product and putting it front and center — in the text between the person and the tool.
The only question worth asking is whether you’ve been paying attention long enough to know what you actually want. I’m still figuring mine out, one rewrite at a time.