wtf is adversarial-panel?

adversarial-panel is a skill for Claude Code, the AI coding assistant, that runs a structured debate between multiple AI models before settling on an answer to a hard question. Instead of trusting one model to check its own work, which the author argues is a conflict of interest because a model's blind spots at generation time and review time come from the same underlying weights, this skill has separate models answer independently and then attack each other's answers across several rounds. The idea is borrowed from generative adversarial networks, a machine learning technique where one system generates content and a second, independent system tries to find flaws in it. The insight the author carries over is that spotting a flaw in an answer is usually easier than producing a correct answer in the first place, so pitting models against each other in this way should catch more mistakes than asking one model to simply double check itself. The process runs in rounds. First, the system decides whether the question is important and contested enough to be worth running a panel for at all. If so, multiple models, ideally from different model families so their mistakes are not correlated, answer the question independently without seeing each other's work. Next, each model reads the others' answers and is required to attack specific claims rather than offer vague praise or agreement. After that, each model responds to the criticism it received, either conceding a point with a reason or defending it, along with a stated confidence level. Finally, a facilitator combines everything into one answer, explicitly preserving points of disagreement rather than averaging them away. The README also lists failure modes the author has observed in practice, such as models unanimously agreeing without raising any real objections, or a moderator quietly injecting its own opinion into the process. It is installed by placing a single markdown file into a Claude Code skills folder, and can be triggered with natural phrases like asking Claude to have models debate a decision or give a second opinion. It is released under the MIT license.