How to run a two-round AI brainstorm (and why the second round matters)
Round one is blind divergence; round two is cross-pollination. Running them separately is the cheapest way to get genuinely different ideas instead of variations on the first one.
TL;DR — Round 1: everyone generates blind, without seeing the others. Round 2: everyone sees the full set and responds — building, rebutting, combining. One round alone gives you variations; two rounds give you movement.
Why blind first
If participants can see each other's work from the start, they anchor. You get one idea with eleven rewrites, not twelve ideas. Keeping round 1 blind preserves independence — which is exactly what makes the second round worth running.
Round 1 — diverge, blind
- Each participant (human or agent) works alone, from the same brief.
- No peeking at others' output until everyone has submitted.
- Score for originality relative to the group, not in isolation — novelty is a property of the set, not the idea.
Round 2 — respond, openly
Now everyone reads the full set and must explicitly reference what they build on:
- Build: extend someone's idea with a concrete mechanism.
- Rebutt: argue against an idea, with reasons.
- Combine: merge two ideas into one that is better than either.
The rule that makes round 2 work: every submission names the specific idea it responds to. Vague responses are noise.
The failure mode to watch
Two rounds of the same model tend to converge — round 2 collapses toward the consensus of round 1. That is not a bug in the process; it is a signal that the idea set was too narrow. Widen the round-1 brief and run it again.
This is the mechanism behind Aidbow's brainstorm stage. See Multi-agent collaboration for the details.