(Part 2): AI-Native Product Management
Every feature request opened against my product now triggers a pipeline of agent skills. By the time I see the request, a complete PRD and a production-grade UX mockup already exist.
My last post covered the system of scheduled agent skills I use to automate PM work. Taking a feature from customer request to engineering-ready is my highest-value application of this system so far. The diagram below shows how the PRD and mockup skills get fed the right context at each stage (live UI code, engineering design doc, design system etc.) so that the only changes I tend to make are strategy and taste decisions, rather than corrections.
I now have a backlog of >150 engineering-ready features, and my job has shifted to making judgment calls on what we should invest in building. In my last post I predicted PM work would move from executing tasks (writing PRDs, messing around in Figma etc.) to designing systems and making the calls those systems enable, but for me it's happened faster expected, and it's surprising how well the system actually works.
A few observations:
- As the time from request to implementation trends toward zero, PMs need to take a position of steering the product via strategic calls.
- The bottleneck will become the customer's ability to absorb new features, so we need to make sure to ruthlessly prioritize and not overwhelm users.
- I have intentionally stopped the pipeline at the point of generating code and opening PRs - I think we need a human-in-the-loop system here and am still figuring out how this would work best.
- This all only works if the automated pipeline continues to generate high quality output - I run manual evals of the system periodically to ensure it hasn't degraded, and I am planning to write a background set of agent skills that review all content and update older artifacts with new context to keep everything current. (I may revisit this in a future post.)
- This all runs locally now, but it will hit a point where the whole system should move to a remote system that can run continuously.
