Most AI-generated marketing plans are forgettable — generic, surface-level, the kind of thing you skim once and never open again. The fix isn’t a better prompt asking for “recommendations.” It’s changing who’s doing the talking.

Instead of asking the model to hand me answers directly, I have it role-play an entire focus group — 100 simulated participants reacting to the company in question. But I don’t let them talk all at once. I break them into 20 small clusters of 5, mixed up at random, and let each cluster debate the company on its own first. Only after that do I merge all 20 conversations together, pulling out the specific, actionable ideas that surface repeatedly or stand out.

What comes out the other end is a real working document — usually somewhere between 5 and 20 pages — followed by a second pass where I compress all of that into a single-page summary anyone can act on.

Here’s a sample 1-page run using Essent as the test subject (no relationship to the company — just a good example to work through). Longer version is also available.