GPT-4.1 Prompting Guide

Overview

• GPT-4.1 models are an upgrade from GPT-4o, with enhanced capabilities in coding, instruction following, and handling long contexts.
• This guide provides prompting tips to maximize the capabilities of GPT-4.1.
• It emphasizes the importance of specific, clear instructions and context provision.
• GPT-4.1 requires prompt migration due to its literal instruction-following behavior.
• Provides examples and advice for effective prompt engineering.

Important Prompting Tips

General Best Practices

• Include context examples and specific instructions.
• Induce planning through prompts to enhance model intelligence.
• Prompt migration may be necessary as GPT-4.1 follows instructions more literally than its predecessors.

Agentic Workflows

• GPT-4.1 excels in agentic workflows, solving 55% of problems on SWE-bench Verified.
• Recommended prompts include reminders for persistence, tool-calling, and optional planning.
• Clear reminders transform GPT-4.1 from a chatbot to an eager agent, increasing task success rates.

System Prompt Reminders

1. Persistence: Ensures the model continues until the task is complete.
2. Tool-calling: Encourages using tools instead of guessing.
3. Planning: Optional; ensures thorough planning before and reflection after each tool call.

Tool Calls

• GPT-4.1 is trained to use tools effectively.
• Tools should be passed through the API’s tools field for accuracy and compatibility.
• Use clear and detailed descriptions for tools and examples of usage.

Prompting-Induced Planning & Chain-of-Thought

• Allows developers to prompt GPT-4.1 to plan and reflect between tasks.
• Inducing explicit planning increased task success rates by 4% in tests.

Sample Prompt for SWE-bench Verified

• Example provided demonstrating agentic task execution with detailed problem-solving strategy.
• Emphasizes understanding problems, codebase investigation, detailed planning, incremental changes, debugging, and comprehensive validation.

Long Context

• GPT-4.1 supports a 1M token input context window, suitable for parsing, re-ranking, and multi-hop reasoning.
• Performance is strong but can degrade with complex reasoning tasks.

Tuning Context Reliance

• Consider the balance of external and internal knowledge.
• Different strategies for using context to answer questions.

Prompt Organization

• Instruction placement impacts performance, especially in long contexts.
• Ideally, instructions should be placed at both the beginning and end of the context.

Chain of Thought

• Encourages step-by-step problem-solving to improve quality.
• Especially effective in agentic reasoning and real-world problem-solving tasks.

Instruction Following

• GPT-4.1 has strong instruction-following capabilities.
• Developers should provide clear and explicit instructions to guide behavior.

Recommended Workflow

1. Start with high-level guidelines.
2. Add specific sections for detailed instructions.
3. Use step-by-step lists to guide model workflows.
4. Iteratively refine instructions based on testing and observations.

Common Failure Modes

• Be aware of adverse effects of rigid instructions.
• Avoid repetitive responses by varying sample phrases.
• Mitigate verbosity by providing specific instructions.

Example Prompt: Customer Service

• Demonstrates best practices for structured customer service prompts with diverse rules, specificity, and examples.

General Advice

Prompt Structure

• Suggested sections: Role and Objective, Instructions, Reasoning Steps, Output Format, Examples, Context, Final Instructions.

Delimiters

• Recommendations for choosing delimiters (Markdown, XML, JSON).
• Special considerations for large document contexts.

Caveats

• Address issues with long repetitive outputs or parallel tool calls.

Appendix: Generating and Applying File Diffs

• Improved diff capabilities in GPT-4.1.
• Recommended diff format example and tool for applying patches.