I agree. There are strategic bottlenecks and technical bottlenecks. But knowing what to build is definitely the first. We actually created a comprehensive skill for this program that analyzes all of the person's chat history on Claude and ChatGPT, looks at what they use most, looks at the value each workflow providees, and then recommends what to build in Claude Code.
That’s an interesting direction essentially turning usage patterns into product strategy.
The key leverage there is correct: what people repeatedly do is often a better signal than what they say they want. If you can reliably surface those patterns, the “what to build” problem gets much clearer before you even touch implementation.
The hard part will be ensuring the analysis doesn’t overfit to recent or noisy behavior, but the core idea is strong: reduce guesswork by grounding decisions in actual workflows.
The interesting assumption is that the bottleneck is learning agents.
For most knowledge workers, the harder problem may be deciding which parts of their work are worth automating in the first place.
I agree. There are strategic bottlenecks and technical bottlenecks. But knowing what to build is definitely the first. We actually created a comprehensive skill for this program that analyzes all of the person's chat history on Claude and ChatGPT, looks at what they use most, looks at the value each workflow providees, and then recommends what to build in Claude Code.
That’s an interesting direction essentially turning usage patterns into product strategy.
The key leverage there is correct: what people repeatedly do is often a better signal than what they say they want. If you can reliably surface those patterns, the “what to build” problem gets much clearer before you even touch implementation.
The hard part will be ensuring the analysis doesn’t overfit to recent or noisy behavior, but the core idea is strong: reduce guesswork by grounding decisions in actual workflows.