LLM Agent Prompt Engineering Tips

1. Restate TODO list to focus attention: When context window is long, LLM loses focus. Put important info at the end and restate TODO list after each subtask. 2. Preserve error content: When retrying, provide the error message description to LLM so it knows what went wrong and can avoid it. 3. Beware of few-shot learning trap: Few-shot examples can homogenize results. For diverse agent tasks, avoid too similar examples. Also, history messages act as few-shot examples; to avoid bias, add noise, use different versions, or avoid similar tasks. 4. Use prompt caching for AI app development. 5. Place important info at the beginning or end of context. 6. Store long content in external files and read when needed. 7. Pre-fill reply content to guide LLM to complete tasks or call/block specific tools. 8. Provide accurate error info for LLM to correct mistakes. 9. Avoid AI being influenced by homogenized history messages.