Karpathy Touts 10-Minute Voice Prompts for Complex LLM Tasks

Large language models can execute complex assignments, but their output often depends on how fully users communicate goals, constraints and context. Andrej Karpathy, a founding member of OpenAI, argues that users may get better results by supplying extensive background upfront instead of repeatedly refining short prompts. The approach relies on an LLM’s ability to reconstruct intent from loosely organized information and turn it into a workable task specification.

In his latest guidance, Karpathy recommended using voice input for difficult tasks and speaking in a stream of consciousness for as long as 10 minutes. Users can describe the problem, relevant history, tentative ideas and desired outcome in one extended briefing. Karpathy said the heavier upfront context can reduce later clarification, repeated corrections and other communication costs, making voice prompting a practical way to collaborate with AI systems.