Note
Build to learn, not to launch
What building an AI MVP solo taught me about scoping agents.
As a founder, I built my AI assistant's MVP myself with Claude Code and Cursor. I'm not an engineer by trade, and that was the point. I wanted to learn fast with real users, not wait for a team.
Building it changed how I work as a PM. I know which features take an afternoon and which take a month. I've seen first-hand where LLMs fail: they lose track of context, sound confident when they're wrong and give different answers to the same input. And I've learned to prototype an idea before writing a document about it.
Now, when I'm scoping an agent, I start with the narrowest version that a real user can try this week. The prototype answers questions that no amount of planning can.