Andrej Karpathy, a former executive at OpenAI and Tesla, has released new advice regarding prompt engineering for large language models (LLMs). Karpathy, who previously coined the term “vibe coding,” suggests that users should stop spending time crafting perfectly polished text prompts. Instead, he recommends utilizing voice mode to ramble continuously at the AI for nearly 10 minutes. According to Karpathy, these long, stream-of-consciousness sessions allow users to provide more information, or "bits," to the model, which helps it better grasp the user's ultimate goal.

Karpathy describes his own approach as a "total mess" where anything goes. He often begins these sessions by warning the model that he is switching to speech recognition, or by turning the session into a short, multi-turn interview. The AI then processes the incoherent monologue to generate structured, clear, and actionable outputs. By dumping all thoughts, nuances, and goals at once, Karpathy claims users improve the "mind meld" with the model, which ultimately requires fewer manual corrections throughout the process. He argues that this method effectively shifts the burden of organization from the human to the machine's processing power.

Andrej Karpathy, a former executive at OpenAI and Tesla, has released new advice regarding prompt engineering for large language models (LLMs). Karpathy, who previously coined the term “vibe coding,” suggests that users should stop spending time crafting perfectly polished text prompts. Instead, he recommends utilizing voice mode to ramble continuously at the AI for nearly 10 minutes. According to Karpathy, these long, stream-of-consciousness sessions allow users to provide more information, or "bits," to the model, which helps it better grasp the user's ultimate goal.

Karpathy describes his own approach as a "total mess" where anything goes. He often begins these sessions by warning the model that he is switching to speech recognition, or by turning the session into a short, multi-turn interview. The AI then processes the incoherent monologue to generate structured, clear, and actionable outputs. By dumping all thoughts, nuances, and goals at once, Karpathy claims users improve the "mind meld" with the model, which ultimately requires fewer manual corrections throughout the process. He argues that this method effectively shifts the burden of organization from the human to the machine's processing power.