Why curation is the job now

AI will give you 50 options before lunch. That is not the win. The win is knowing which one to keep.

I have been writing that AI has no taste — models average what they have seen, so the first draft always looks fine and a little anonymous. What changed in 2026 is the math around that problem. The generation step got so cheap that the curation step became the whole job. If you are a solo founder, you are not a generator anymore. You are an editor with very expensive taste and very cheap tools.

What does a 29% hit rate mean?

It means even a good eye throws away most of what AI makes.

A solo studio in Berlin shared his numbers this spring: nearly 20 years of design experience, six digital products, 98 blog articles, and 116 merch items all run by one person. He generates visuals with AI tools daily — then filters hard. Out of about 400 raw AI concepts, 116 made it into the store. That is a 29% keep rate, and he calls himself generous.

That number is the point. He is not failing at prompting. He is doing the job correctly. AI gives you volume. Your eye decides what survives contact with a real product.

You see the same ratio in component work. One founder I talked to rebuilt every UI piece after her first Figma Make pass felt like AI slop — eight to ten options per component before she locked one. Cards, accordions, priority views: four layouts on the table, one ship. That is not indecision. That is curation with receipts.

Why is generation now the cheap part?

Because every tool already won the speed race.

Figma's 2026 survey of product designers tracked the average stack jumping from three tools to seven in a year. Cursor, Claude Code, Lovable, v0, Galileo — they all generate fast now. Anthropic's guide to building evals for agents says it plainly for probabilistic systems: you score the output against what you decided was good, you do not grade the prompt.

Speed is solved. Judgment is not.

When everyone can generate a decent landing page in an afternoon, "can you ship?" stops being the filter. "Did you cut enough?" becomes the filter. The products that stand out in 2026 are not the ones with more screens. They are the ones that kept fewer screens and made them feel like they came from the same person.

Stripe's analysis of top solo founder traits found the same gap at the business layer — the top 10% of solo founders now earn at a 61x gap over the median. They do not use more tools. They keep more users. And keeping users is a curation problem: fewer features, held to a higher bar, shipped consistently.

How do you build a curation system that ships?

You make generation parallel and curation singular.

Generation should run wide. Spin nine terminals, try four models, let Codex draft while Claude critiques — that part is cheap. Curation should run narrow: one set of components, one written rubric, one person who says yes or no.

Here is the loop that works for me:

Lock the set before you scale it. Pick your buttons, radii, type scale, and spacing once. Six buttons, not sixteen. One typeface pairing, not three. When the set is locked, AI output has something to fail against. Without it, everything looks "fine."

Write down what good means. Not in your head. On a page. Three checks: does it match intent, does it stay consistent with the set, does it handle an edge state without falling apart? That is the same rubric I use when your AI eval set is your taste, made measurable — a short list beats a long feeling.

Review on a schedule, not by mood. One Berlin founder runs a full audit across 100+ screenshots every morning — desktop, tablet, mobile — scored from rough to professional. I do a lighter version: ship window in the morning, review window right after. Batch the cuts. Do not curate in the same session you prompted. Your brain needs distance to see the slop.

Cut at 70%. If you generated ten options, pick two or three. If you kept seven, you did not curate. Archive the rest where you cannot see them.

When does curation become taste?

When you do it enough times that other people can describe your call.

Taste is just curation repeated until it forms a pattern. Early on, it feels like personal preference — "this one feels warmer." After 50 cuts, you can name the rule: warmer meant looser line height, softer shadows, fewer transitions. After 200 cuts, someone else can apply the rule and get close to your pick.

That is why the solo founders I keep watching do not chase new models every week. They chase a tighter rubric. They know fewer tools will make you ship faster not because the tools are bad, but because each new generator adds more to curate and no new way to decide.

My take is simple: do not try to get better at prompting this month. Get better at saying no. Generate less, filter harder, keep the bar visible on a page you open every morning. The model will not learn your taste between sessions if you do not write it down. But you will — one cut at a time, until curation is the product and generation is just the supply chain.