How do you make several AI personas actually reason differently?
I set up a few personas, like a skeptic, an optimist and an engineer, to discuss a decision. They end up with different voices but the same conclusions. How do you get personas that genuinely think differently instead of just sounding different?
Personas that share one chat see each other's words, so they drift toward agreement. The first fix is isolation: let each persona write its view separately, without seeing the others, and only bring the answers together afterwards.
A name and a tone are not a way of thinking. Give each persona a different job and different evidence: one judges cost, one hunts for failure cases, one argues for the user. If they all get the same brief and the same goal, they will reach the same answer in different words.
Make disagreement part of the format. Ask each one to state what it would need to see to change its mind, then run a round where each must answer the strongest point against it. A final step should weigh the views, not average them.
If the personas still agree, the question may simply have one obvious answer. That is a useful result too, not a prompt failure.
Listings mentioned
- Council · skill by danielmiesslerRuns topic-briefed agents through a visible multi-round debate built for real friction.
- /cs:boardroom — Multi-Role Boardroom Deliberation · skill by alirezarezvaniKeeps roles isolated in an early phase, screens with a critic, then writes one synthesis memo.
Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.