According to the Watchful AI Monitor, OpenAI co-founder and current Anthropic pre-training team member Andrej Karpathy shared a "Long-Form Oral Prompt" working method. Faced with a complex task, he would open up voice input and speak continuously for about 10 minutes. The content can be fragmented, chaotic, and does not need to be organized into a complete instruction beforehand.
Karpathy believes that large models excel at reconstructing real objectives from these pieces. They can also organize tangled thoughts more clearly. When necessary, he would have the model ask follow-up questions for a few rounds, turning the oral narration into a small interview. He would provide plenty of background information in the early stages, leading to fewer corrections needed later on.
