The author recently posted a prompt on Large Language Models (LLMs) like Claude sonnet, GPT-4o, and Deepseek v3, instructing them to contemplate before providing final answers, which unexpectedly went viral. The inspiration behind the prompt was exploring reasoning skills, following OpenAI’s latest models. The author discusses the main principles of building the prompt, emphasizing exploration over conclusions, depth of reasoning, and realistic thought processes. They also outline the importance of persistence and separating contemplative phases from final answers. While this contemplation phase may help with reasoning, LLMs are susceptible to hallucinations and may not always guarantee correct responses, making it more suitable for intermediate to difficult tasks.
https://maharshi.bearblog.dev/contemplative-llms-prompt/