🧠 The “double pass” phenomenon: how to make an LLM smarter with a simple copy-paste
While developers spend weeks perfecting their prompt engineering, practice shows that sometimes the most effective solution is as simple as it gets.
What’s the hack?
Studies confirm that if you send the same prompt twice in one message (just duplicate the text, one copy right after the other), the model produces significantly better analysis.
Why does it work?
Neural networks normally work in a linear way. Each new token “sees” only what came before it. When you hand over a huge amount of data and put the question at the end, the first chunks of context get processed without a clear understanding of the final goal.
During the second pass (the duplicate), every segment of the text already “knows” where you’re going with it. This creates an internal feedback effect.
• Results: In tests on finding specific information in a large dataset, accuracy rose from 21% to 97%.
• Speed: There’s almost no delay, because modern hardware processes the input text in parallel.
To minimise the model’s hallucinations on complex analytical tasks, add this instruction at the end of your prompt:
“Before giving your final answer, list 3 arguments against your own solution and refute them.”
This forces the neural network to go beyond its first linear hypothesis and run an internal check. Combined with duplicating the prompt, it gives you the cleanest, most professional result possible.
