Notes6 min read
Building with AI got cheap. Judgment didn't.
Judgment. Deciding what deserves to exist, knowing what good looks like, and getting it in front of people who will pay. Agents can do more and more of the building, but those calls are still yours.
Not long ago, the hard part of making software was the making. You needed engineers, months and a budget. Most ideas died waiting for all three, which meant the ones that survived had at least been argued over.
That wall has mostly come down. With an agent like Claude Code, a designer or a PM can describe a small app in the morning and click through a working version after lunch. We think that's good news. It also changes the job. Once AI can build almost anything, the hard part is judgment.
The hard part once AI can build anything
When something gets cheap, people make more of it. Stripe can see this in its own data. In the second quarter of 2026, 63% of the companies incorporated through Stripe Atlas had a single founder, the highest share it has recorded. Stripe's explanation is that AI tools make it easier for one person to build, ship, support customers and keep improving.
The same report has a second finding that gets less attention. For solo-founded startups, median revenue in the first six months fell 23% in 2025, while the top tenth grew 19%. Four years earlier, the top tenth made 34 times what the median founder made. Now it's 61 times.
So more people are building, the typical result is getting worse, and the distance between typical and best keeps growing. We read that as a sign of where the work went. Generating a product is now the cheap, crowded middle of the journey. The scarce part moved to both ends: before you build, deciding what deserves to exist, and after, making it good and getting it to people who pay.
There's a second reason the middle feels crowded. AI models lean toward the most common answer. Anthropic calls this "distributional convergence." Safe design choices dominate the web that models learned from, so without direction a model drifts to that safe center. Figma's CEO Dylan Field makes the same point: the first thing AI gives you is generic, because it's an average of everything it has seen. So cheap building gives us more products, and more of them look alike.
Three judgments that don't get cheaper
By judgment we mean making a call when there's no right answer to look up, and being able to explain it afterward. Three of them matter most, and each sits at a different point in the journey.
What to build
Paul Graham has long argued that the most common startup mistake is solving a problem nobody has. Cheap building makes each wrong guess cheaper, but it also lets you make ten wrong guesses a month. His advice is to notice ideas rather than invent them: live in the future, then build what's missing. Ideas you sit down and make up tend to sound plausible and go nowhere.
Agents help a lot here. They can scan forums, summarize interviews and sort a hundred complaints by theme in minutes. What they can't do is care which problem deserves your next year.
Take Kudos, a made-up app. It helps freelancers, coaches and small businesses collect testimonials from their customers. An agent can build the request form in ten minutes. Whether coaches actually struggle to ask for testimonials, or only say they do, is something you learn by talking to coaches. Writing the press release first is a cheap way to test an idea before you build it.
What good looks like
This is taste: knowing what's good and being able to say why. Graham makes a useful argument here. If taste were only personal preference, it couldn't get better. But designers' taste does improve as they grow, so their earlier taste was worse, not just different. His recipe for great work is exacting taste plus the ability to satisfy it.
AI has handed almost everyone the second half of that recipe. You can generate twenty versions of the Kudos testimonial wall before your coffee cools. The first half is still on you: looking at those twenty, knowing which one has a point of view, and saying why the other nineteen don't. We wrote more about this in the taste gap.
Who it's for and how they find it
A product nobody hears about is a hobby, which is why AI-built apps get no customers. Agents can draft your positioning, research prospects, write the follow-up email and send it once you approve. Choosing who you're for, and what you'll say to them, is a judgment about people.
It's also a judgment about limits. Kudos collects testimonials, and since 2024 the FTC has banned fake ones, including AI-generated reviews and reviews from people who never used the product. An agent will happily invent a delighted customer if you let it. Saying no stays with you.
| Judgment | What agents can do | What stays with you |
|---|---|---|
| What to build | Scan forums, summarize interviews, score ideas | Choosing the problem worth a year of your time |
| What good looks like | Generate twenty variants, follow your written rules | Picking the one with a point of view, and saying why |
| Who it's for | Research prospects, draft outreach, send after approval | Choosing the customer, and what you will never fake |
What this means for how you learn
If building is only part of the job, a course that ends at "it's deployed" teaches only part of the job. That's why Chief of Agents follows one product through five stages, and Build is just one of them.
- Foundations. Set up Claude Code, learn the words in plain language, meet your crew of agents, and start training your eye in Taste 101.
- Ideate. Notice a future, find a real problem in it, talk to customers, write the press release before you build, and decide go or no-go.
- Build. Write your taste down as a DESIGN.md, plan, build, review and ship with payments connected.
- Sell. Position the product, launch it, make your own creative with AI, and work toward a first paying customer, with real testimonials only.
- Run. Hand routine work to scheduled agents so the business keeps going without you watching it all day.
Then there's Demo Day, where you show what you made and who's using it.
Claude Code's features show up along the way: skills, hooks, scheduled routines, agents working in parallel. But each one arrives when the project needs it, as a way to carry a decision further, not as a feature tour. The lesson is always the decision. The feature is how you hand it off.
Where that leaves us
We don't think the people who do well in this period will be the ones who generate the most. Generation is the part everyone has now. The ones who do well will choose carefully, throw away most of what the agents make, and stay close to the people they're building for.
Our agents can do the work. Deciding which work is worth doing is still ours.
Sources
- Solo founding is at an all-time high: Top performers have these traits in common (Stripe, May 2026)
- Improving frontend design through Skills (Anthropic, November 2025)
- Figma CEO on how anyone can get good at design in AI (Behind the Craft, April 2026)
- How to Get Startup Ideas (Paul Graham, 2012)
- Taste for Makers (Paul Graham, 2002)
- FTC announces final rule banning fake reviews and testimonials (Federal Trade Commission, August 2024)
Learn this inside Claude Code.
Join the waitlist for Chief of Agents. You'll get the free beginner's handbook and skills right away, and hear first when the course opens.
Keep reading
6 min read
From designer to builder: directing a crew of agents at work
What working with AI agents looks like day to day after ten years in design: the work of direction, and what stayed the same.
6 min read
AI closed the skill gap. It widened the taste gap.
Why does AI output look generic? It drifts to the average. Here's why, plus a side-by-side drill that trains your eye to fix it.
6 min read
Everyone's vibe-coding. Nobody's getting customers.
Why do AI-built apps get no users? Shipping got easy, but finding paying customers didn't, so the course spends two stages on it.