Most of us end up with a folder of product context to feed the AI: strategy, decisions, metrics, PRD. Mine worked great for about two weeks. Then the team shipped, the folder didn’t move, and the AI was confidently answering about a product that no longer existed.
The four easy loops you can keep by hand with discipline: feedback clustering, decisions, metrics, discovery. The fifth one I could not: “the team shipped, so update the product state”. Because the trigger is really “someone marked a task done”, and done is an opinion, not proof the code does what the story asked.
What changed it for me was moving the trigger to verification against the criterion, not a click, and tying each piece of code to the artifact that asked for it so the brain updates itself.
How are you keeping yours from going stale? Manual ritual, or something that feeds it on its own?
I wrote up the full argument here: The product brain that maintains itself · PAELLADOC