Paste Your Prompt TWICE: Google's Free Trick to Make AI Smarter (New Paper)

What is the One-Line Trick to Improve AI Responses?

Introduction to the Trick

• The trick involves pasting your question twice in a single message, which can lead to significantly better answers from AI models without any additional cost or complexity.

• This technique is based on research published by Google researchers Leviathan, Kalman, and Matias in December 2025.

Research Findings

• The paper titled "Prompt repetition improves non-reasoning LLMs" states that repeating prompts enhances performance for popular models like Gemini, GPT, Claude, and DeepSeek.

• The method is straightforward: simply copy your query and paste it again immediately after the first instance.

Mechanism Behind the Trick

• Language models read input left to right; thus, when a prompt is repeated, the second instance allows each word to reference the entire first copy.

• This change enables better context understanding as options can now see the question they are answering.

Effectiveness of Prompt Repetition

• In tests across seven popular models on various benchmarks, prompt repetition won 47 out of 70 tests with zero losses.

• A notable improvement was observed where one model's accuracy jumped from 21% to 97% on a recall task due solely to this technique.

Cost Implications

• Sending prompts twice does not significantly increase costs or latency since reading inputs is fast and cheap compared to generating responses.

• The paper confirms no increase in generated tokens or real added latency when using this method.

Limitations of the Technique

• This trick only benefits non-reasoning tasks; reasoning models do not gain from repeated prompts as they already process questions effectively on their own.

• It’s best suited for quick factual inquiries rather than complex reasoning tasks.

Practical Application of the Trick

• To implement this strategy: take your prompt, copy it directly below itself, and send.

• Be cautious with long prompts; doubling them may exceed limits or slow down processing times.

Additional Tips for Effective Prompting

• Place crucial instructions at either end of your prompt for maximum impact.

• Keep prompts concise; excessive information can degrade model performance.

• Experiment with this technique across different models and tasks for optimal results.

Conclusion & Resources

• For further learning, resources including a cheat sheet summarizing these techniques are available upon request through comments.