Squad Communication

Information flows sequentially from lead through specialists, with a feedback loop between generator and critic at the end.

The lead sends requirements to the researcher, who sends a style brief to the prompt architect, who sends five prompts to the generator. The generator produces images and passes them to the critic. The critic either approves (scores are above threshold) or sends improvement instructions back to the generator for another round. This generate-critique loop runs up to three times. The pipeline is linear except for the feedback loop at the end — that's where iteration lives.

The generator-critic loop runs up to three rounds. After that, diminishing returns kick in — take the best result and move on.

PaperBanana research showed accuracy gains plateau after three critique rounds. Round 1 catches obvious issues — wrong composition, missing elements, color mismatches. Round 2 catches subtleties — text legibility, spacing imbalances, tonal inconsistencies. Round 3 polishes — fine details, visual harmony, professional finish. Beyond three rounds, the improvements are marginal and the token cost is real. Set a hard limit and take the critic's top-ranked result.