Andrej Karpathy, the prominent artificial intelligence (AI) researcher who famously coined the term “vibe coding,” has a piece of advice for anyone spending hours crafting the “perfect” text prompt for AI: stop trying so hard. Instead of typing out neat, polished instructions, Karpathy suggests users of AI chatbots should simply open the microphone and ramble.In a post on X, the former OpenAI and Tesla executive, who is now a leading AI researcher at Anthropic, revealed that his go-to prompting strategy involves turning on voice mode and talking continuously at a large language model (LLM) for nearly 10 minutes at a time.“Sometimes the LLM needs more bits to understand what you’re trying to achieve,” Karpathy wrote, describing his rambling voice sessions as a “total mess, anything goes, full stream of consciousness.”
‘Let AI do the organising’
According to Karpathy, rambling aloud creates a far better division of labour between human intuition and machine processing power. He said that speaking allows humans to quickly dump every half-formed thought, nuance and goal without getting stuck on phrasing or grammar.The AI takes the chaotic monologue and organises it into clear, actionable and structured outputs. He added that by providing maximum context upfront, users “improve the mind meld and have to correct things less from that point on.”
Read Karpathy’s full post on X (formerly Twitter):
One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you’re trying to achieve, but you’re too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like “switching to speech recognition sorry for any typos...”. Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.