Why your agent writes generic PRDs

When it all started, I asked an agent to write a PRD for me. It came back so general it missed every nuance. Very meh. I read it twice looking for one sentence that could only describe my product. Zero.

It failed in the most boring way possible. No hallucinations, no drama. Just a spec-shaped object with nothing inside.

The agent wasn’t dumb. It was starving. It had no map of my context, so it guessed.

So I started writing things down. One file: how I write. Then another: what I’m working on. Then the decisions I’d already made, the users I serve, the constraints that are real. Plain markdown, in a repo I own.

Nothing dramatic happened in week one. But every interaction wrote back into those files. A correction here, a preference there. Week after week, the context got thicker and the outputs got sharper. Compound interest, but for context.

Now the same files run on five different assistants. I take my context with me everywhere. No single vendor holds it.

Context is not a prompt you freeze. It’s a living thing. It needs feedback loops. Every conversation should make it smarter, or it’s just another document going stale.

Users mapped to research notes. Constraints mapped to a decisions file. Less context, but mapped, and alive.

#employablePM