The creator expected to prove AI could match Frost and Poe, but users had zero interest in poetry.
The Summary
- Two builders launched PoemAIGenerator.com expecting poetry demand, but discovered people wanted something completely different from their LLM tool
- Simple calculator sites routinely generate $10,000+ monthly despite being just lines of code, proving utility beats creativity in traffic arbitrage
- The gap between what builders think users want and what users actually need reveals the real opportunity in agent-powered tools
The Signal
Two developers built a poem generator to prove LLMs could master creative writing. They vibe-coded the interface in under an hour using ChatGPT, hooked it to OpenAI's Assistants platform, and launched. The hypothesis was simple: poetry requires emotion and lived experience, therefore an AI that could write decent poems would prove something meaningful about machine creativity.
The article cuts off before revealing what users actually did with the tool, but that cliffhanger itself is the signal. The builders went in with a theory about AI capabilities and came out learning about human behavior instead.
"Calculator sites are built around just a few lines of code but generate ungodly amounts of traffic and revenue."
This is the Web4 pattern emerging everywhere. The technology is table stakes now. You can spin up an LLM-powered tool in an hour. The hard part is not building the agent, it's understanding what job people actually hire it to do. These developers thought they were in the poetry business. Based on the article's framing, users had other plans.
The calculator site economics here matter. These builders studied sites earning five figures monthly from simple utilities: character counters, title case converters, mortgage estimators. The traffic comes from solving tiny, repetitive problems that people Google constantly. Add an LLM and suddenly your calculator can handle fuzzier inputs and generate paragraphs instead of numbers.
Key shift happening:
- Old web: Build a calculator for one narrow task, hope for search traffic
- Agent web: Build one LLM interface, let users define the task
- The winner: Whoever figures out what job people keep trying to hire the tool for
The most interesting companies in the agent economy right now are not the ones with the best technology. They are the ones paying attention to how people actually use the tools once they are released. Product-market fit used to mean building what users said they wanted. In Web4, it means watching what they do with your agent when you are not looking, then rebuilding around that behavior.
The Implication
If you are building with LLMs, ship fast and watch usage patterns obsessively. The gap between your hypothesis and user behavior is where the actual product lives. These developers learned more from unexpected usage than they would have from six months of planning the perfect poetry tool.
For workers, this is the pattern to watch. The valuable skill is not coding the agent. It's seeing the job-to-be-done that everyone else missed, then pointing the agent at it. The developers who get rich in Web4 will not be the best prompt engineers. They will be the best human behavior translators.