Karl's AI Watts proposes three-stage AI reasoning-level workflow

Karl's AI Watts (@aiwarts) describes a three-stage workflow for working with AI models, which he calls his "big AI theory." In stage one, the user sets the model to the highest reasoning level, selects plan mode, and gives the model voice, image, and video input instead of typed prompts.

In stage two, the user drops to medium reasoning to carry out the work from that plan. In stage three, the user returns to high or higher reasoning to check the completed work.

Karl says a detailed spec run at medium reasoning beats a vague request run at high reasoning on Terminal Bench 3.0, and that the final review step catches an extra 20%-30% of problems. These figures are his claims in the post, and the report does not include a published benchmark or method behind them.