Why this works (in plain English)

New tools, long prompts, better copywriting. Same decisions, same bottlenecks.

Here's a trick that quietly cuts through the noise:

Before you spend 30 minutes crafting the "perfect" prompt, try sending the exact same prompt twice in one message.

Literally:

[your prompt][your prompt again].

Google researchers tested this across seven mainstream models (Gemini, GPT, Claude, DeepSeek) and seven benchmarks.

Prompt repetition improved accuracy in 47 of 70 tests, with zero losses.

On one "find the right name in a long list" task, one model jumped from 21% to 97% accuracy just from duplication. Not a new model. Not a new feature. Just copypaste.

LLMs read left to right. Early tokens can only "see" what came before them.

So when you write:

"Here's three pages of context...Question at the bottom."

The question sees the context, but the context never saw the question. The model is not factually wrong, just contextually wrong.

When you repeat the full prompt, every token gets a second pass where it can attend to every other token.

You're effectively letting the model read the brief twice before answering, without changing the output length or noticeably slowing it down in most normalsized prompts.

It's not miraculous. Real.

When to use this in your business

This works best on nonreasoning tasks where the model is pulling something specific out of messy text:

• Finding a name, ID or code in a long list or email thread.
• Classifying tickets, complaints, or support emails.
• Extracting key fields from contracts, reports or meeting notes.

If you're already using a "reasoning" mode or detailed chainofthought prompts, the gains drop off – those models already reread the prompt internally.

Nextweek activation: the 30minute "prompt repetition" test

Pick one workflow where your team already leans on an LLM but doesn't fully trust the answers.

Customer emails, finance checks, document search – whatever hurts most.

Then run this miniexperiment:

• Take 20–30 real examples from your business.
• Run them once with your current prompt, log accuracy.
• Run the same examples again, but with the prompt duplicated in the same message.
• Compare: did accuracy, cycle time, or rework improve?
• If you see a meaningful lift, bake "prompt 2" into that workflow and write it into the playbook.

No new tools. No extra dashboards. Move a metric, not a use case.

Start with accuracy or rework rate; if it shifts, you've just increased quality for the price of copypaste.

If you want help turning tricks like this into proper operating logic – where AI consistently helps your business sense–think–decide–act – that's the conversation I'm having with NZ leadership teams every week.