How I Used 6 LLMs to Write One Word Doc Comment

One comment. Six frontier models. Five rounds. I shipped one framing out of ten.

I was reviewing a technical document and wanted to add one precise comment — a specific risk vector that the existing analysis had missed. Not a groundbreaking insight. One bullet point, three sentences, two citations. I ran it through six models in parallel — council mode for consensus, red-team mode to attack each angle — five times over.

The thing that made the final version good was not what the models generated. It was what I threw away. Nine framings sounded credible. They used the right terminology. They cited the right sources. They would have survived a casual read. But they were not sharp enough. They restated known positions or added nuance nobody had asked for. The one that survived was specific — a concrete test vector, not a general concern — and it matched the document’s existing structure.

This is the part that surprised me about multi-model deliberation. The value is not six perspectives. It is six chances to be wrong, which means you can set a much higher bar for what counts as right. When you have one draft, you ship it because you have nothing better. When you have ten, you can afford to ask whether any of them actually deserve to exist.

The models are not good at this filtering step. They will tell you all ten are strong. They will suggest combining three of them. They will not tell you that nine should die. That requires a different kind of judgment — not “is this well-written?” but “does this earn its place in the conversation?” The human’s job in a multi-model workflow is not writing. It is killing.

I would do it again for something this small. The confidence you get from knowing that your one contribution survived adversarial review from six frontier models is worth more than the time it took. When a single comment can shape how people perceive your judgment, the bar for “good enough” should be higher than any individual model can set.