My friend spent two weeks writing acceptance criteria for an AI summarisation feature. By the time it was signed off, the team had shipped three prompt variants, the vendor had updated the model, and the criteria described a feature that no longer existed.
Traditional product specs (PRD) assume everything behaves predictably. With AI, nothing is guaranteed. It shifts, drifts, and surprises us.
Now the question is: If PRDs don’t hold up, what replaces them?
I wrote an article recently about the importance of Evals.
Here is the link:
https://medium.com/analysts-corner/stop-writing-prds-for-ai-features-instead-start-writing-evals-a4bab4f09484
Top comments (0)