GPT 4.1 Prompting Guide

Prompting Tips for GPT 4.1

1. Follow instructions literally. GPT-4.1 is trained to follow directions more precisely than previous models. Be explicit about what you want.
2. Place instructions strategically. For long context, put critical instructions at both the beginning AND end of your prompt for best results.
3. Use specific delimiters. Markdown headings, XML tags, and backticks help the model understand structure. JSON performs poorly for document collections.
4. Induce planning with prompting. Ask the model to "think step by step" when solving complex problems to significantly improve accuracy.
5. Design agentic workflows with clear reminders:
   • "Keep going until the problem is completely resolved"
   • "Use tools when uncertain instead of guessing"
   • "Plan extensively before each action"
6. Leverage the 1M token context window wisely. Performance stays strong to the limit, but degrades when retrieving many items or reasoning across the entire context.
7. Balance internal vs. external knowledge. For factual queries, instruct the model to "only use provided context" or "combine context with basic knowledge" based on your needs.
8. Format your prompts with clear sections:
   • Role and Objective
   • Instructions (with subcategories)
   • Reasoning Steps
   • Output Format
   • Examples
   • Final Instructions
9. Guide information retrieval. When working with documents, ask the model to first analyze which ones are relevant before attempting to answer.
10. Avoid rare prompt patterns. The model may struggle with extremely repetitive outputs or parallel tool calls. Test these cases carefully.
11. Be direct with corrections. If model behavior is unexpected, a single clear sentence is usually enough to steer it in the right direction.
12. Use specific frameworks for coding. For generating code changes, use the V4A diff format with context lines for maximum accuracy.
13. Remember it’s not a reasoning model. GPT-4.1 doesn’t automatically provide an internal chain of thought, but you can explicitly request it to show its work.