Three System Prompt Lines That Boosted OpenAI's Coding Benchmark by 20%

OpenAI published the three instructions their engineers add to every system prompt for AI agents. When tested on the SWE-Bench coding benchmark, these three lines raised the pass rate by nearly 20%. The lines are simple, and you can add them to your own prompts today.

The Three Lines

1. Persistence. Tell the model to keep going until the job is completely solved before ending its turn. This stops the model from handing control back to the user too early, which is a common failure mode in agent workflows. Without this instruction, models tend to partially solve a problem and then ask for confirmation.

2. Use your tools, do not guess. If the model is unsure about code or file contents, it should open them and verify. Do not hallucinate. This single line cuts wrong answers significantly because it forces the model to ground its responses in actual data rather than generating plausible-sounding guesses.

3. Plan, then reflect. Plan thoroughly before every tool call and reflect on the outcome after. This forces step-by-step thinking instead of blind trial and error. The model reasons about what it expects to happen before taking action, then checks whether the result matches.

The Results

OpenAI tested these three reminders on the SWE-Bench coding benchmark. The pass rate jumped by nearly 20%. For something that takes about 30 seconds to add to a prompt, that is an enormous gain. The takeaway is that the model already has the capability; it just needs explicit reminders about how to use it.

Key Takeaways

• Adding "keep going until done" prevents the model from bailing out mid-task
• Telling the model to use tools instead of guessing reduces hallucination
• Requiring a plan-then-reflect loop forces structured reasoning and catches errors early
• These three instructions together raised SWE-Bench pass rates by ~20%

Resources

• OpenAI GPT-4.1 Prompting Guide -- the source documentation for these system prompt recommendations
• SWE-Bench -- the coding benchmark used to measure the improvement