Everyone Is Creating With AI

When a tool makes creation cheap for everyone, it stops being an advantage for anyone. The catch isn’t hidden in the technology, it’s in what happens to value once the bottleneck moves. Here are the five catches, and the skills that get you past them.

The distribution catch

Scarce before: output → scarce now: attention and trust

Creating has never been easier; getting anyone to care is harder than ever. When anyone can generate an app, a newsletter, a logo, or a song in an afternoon, the bottleneck shifts from making to distribution, taste, and trust. The scarce thing is no longer the output.

The depth catch

AI gets you 70 to 80% → the value lives in the last 20%

AI gets you most of the way fast. The last stretch, the part that makes something actually good, reliable, or differentiated, still takes real expertise. Plenty of AI-built products demo well and fall apart under real users, edge cases, or scale. Skip the fundamentals and you can’t debug or extend what you “built.”

The sameness catch

Same models + same prompts → convergent output

Tools trained on the same data, used with similar prompts, produce similar things. If your product could be regenerated by anyone with the same prompt, you don’t own an advantage. The moat has to come from somewhere else: your data, your audience, your judgment, your specific knowledge of the problem.

The economics catch

Value flows to models and infrastructure → not to thin wrappers

Most of the money is pooling at the model and infrastructure layer, not with the people building thin layers on top. A lot of “AI businesses” are one API price change, or one new model feature, away from being obsolete.

The survivorship catch

Visible: the launches → invisible: the graveyard

You mostly see the wins. The feeds are full of “I built this in a weekend” posts, and you don’t see the large majority that got zero users or quietly died. The sample you’re watching is pre-filtered for success.

Metaskills for the last 10%

AI closes the first 90% faster every year. The gap that remains is closed by metaskills: capabilities that sit above any single tool, transfer across every project, and compound instead of depreciating when the models improve.

Taste. Knowing what good looks like. AI produces plausible by default; you have to recognize excellent, reject the rest, and know the difference on sight. Taste is what turns the 90% draft into the thing people actually want.

Decomposition. Breaking a vague ambition into small, checkable pieces the tools can execute well. Output quality tracks framing quality, so the person who structures the problem controls the result.

Specification. Saying precisely what you want. Prompting is just specification, and writing a clear spec, with constraints, examples, and what “done” means, is a durable skill that predates AI and will outlast any model.

Verification. Reading output skeptically and testing before shipping. Finding the failure cases before your users do is the difference between a demo and a product.

Debugging. Reasoning about why something broke when the model can’t tell you. This is where fundamentals pay rent: enough grounding in how the thing works to trace a failure instead of regenerating and hoping.

Domain depth. Understanding the real problem well enough that edge cases are familiar, not surprises. The model knows the general case; you have to know your specific one.

The upshot

None of this means don’t build. It means the leverage is real, but the easy part isn’t where the value is. The people winning combine AI’s speed with something AI can’t supply: a real problem they understand deeply, an audience that trusts them, or taste.

AI speed × (your problem + your audience + your judgment) = the part that’s still yours.

---

October 2026. Subject to revision as the bottleneck moves again.