According to Beating, Andrej Karpathy, OpenAI co-founder and current member of Anthropic's pretraining team, has shared a voice-based prompt technique. Instead of crafting written instructions, users open voice input and speak continuously for around 10 minutes about a task, with no need for coherent or organized thoughts.
Karpathy noted that large language models excel at extracting the actual goal from these fragmented utterances and can organize unclear thinking more clearly. He often follows up with multiple rounds of questions to turn the initial monologue into a mini-interview, finding that providing more context upfront leads to fewer corrections needed later.