There was a conversation I needed to have with another program manager, and I kept not having it. He didn't report to me, he was good at the job, and I didn't want to spend the working relationship badly. So I practiced it first, against a Gem built to play him. It pushed back the way he pushes back, told me exactly where I was hedging, and by the time I sat down with the real person, nothing in it was a surprise. Full prompt at the bottom of the post.

I used to send my status report after one AI pass and my own read-through. Now five personas argue about it first, the business partner, the executive, the chief of staff, and two more, each built around someone who actually reads the thing. This isn't really about status reports. It's a structure you can point at any output you produce.

I built a Gem in twenty minutes to handle a framing problem I'd been fighting for months. That part worked. What I didn't expect was what happened when I started giving it stakeholder feedback — and asking it to improve its own instructions.