I asked Claude 200 times for a zero to one idea. It failed each time.

I spent a day trying to find that moment through Claude. Here is what happened.

Large Language Models such as Claude have an extraordinary ability to remix and recombine. They often return the most sophisticated remix of everything that already exists. The remix is dressed up in novel, seemingly profound language, and delivered with unimpeachable confidence.

Because Claude is built from what already exists, it cannot see what doesn't exist. I asked Claude for zero to one ideas and I repeatedly got zero to 0.3.

Take Noonface, a tool where a founder posts their product, the tool then finds live conversations where buyers are expressing pain in relation to what the product solves, then it writes an entry post. The idea sounded novel. I asked Grok to approximate it with a prompt. Grok provided free links to live conversations where buyers expressed pain in an instant. The idea had it's legs chopped off before it could crawl. Dead.

Killorbuild was meant to be a live-or-die market for software ideas. If you believed something should be built you parked £20 on it, your word is your bond. Builders who claimed then abandoned the idea went on a permanent shame ledger. A verification agent would check whether anything shipped. In theory it was a beautiful design. It died as soon as it was clear it needed three sides at once. Bonders to park money on an empty board, builders to take a public ship date for a tip jar. Each side waits for the other to go first.

Candour, a semi-anonymous whistleblowing channel for professionals to surface what their institution needs to hear but nobody feels safe to say. A slingshot of GDPR, employment law, and NHS DSP Toolkit compliance killed the monstrous idea in one shot. 

Zero to one ideas usually come from a particular person standing in a specific place noticing something nobody else has noticed from that angle.

Gates saw software licensing before anyone else because of where he stood. Zuckerberg saw the social graph because of his particular experience. @jonathan_wilke built outbid.lol because he saw that attention for sale works differently in the AI age, where building has never been easier. 

Claude processes more documented human pain than any individual human can accumulate in a lifetime. The data lives in clinical literature, Reddit threads, X posts, blogs etc. However, data cannot beat the frustration of a person who encountered the same broken thing for the twentieth time and finally decided to fix it.

The zero to one moment comes from the person who can't find the product they needed and built it instead. An LLM that was asked to imagine what that product might be cannot match that experience.

Claude is only useful at killing ideas faster.

That result of the day was a filter. Three questions killed everything I put through them.

Can a sophisticated user approximate this with a single well-crafted prompt? If yes, dead.

Does this require three parties to show up simultaneously before value exists? If yes, dead.

Do compliance requirenents exceed what a solo founder can tolerate? If yes, dead.

You can run the same three against whatever you are building and they might apply to your own projects.

Stop asking what to build and start noticing what's slightly wrong, in a defined, specific sense. Something you encounter and think why has nobody fixed this properly. A pain that has been normalised, so familiar you have stopped noticing it.