Confirm the lost-in-the-middle effect in your own prompt

My long prompt has the right fact in it, and the model still misses the one buried in the middle.

Beginner to intermediateRead 4 minDoing: about an hour7 steps

▶ Start

What you'll have at the end

Three saved answers to the same question, with the key fact moved to the start, middle, and end

Before you start

You need

A long prompt you already send routinely (a system prompt, a pasted document, several reference passages), with the ability to send at least three separate, history-free calls to the model it runs against.

Not covered

Fixing the bias once you've confirmed it, by reordering chunks, cutting the document down, or reranking, is a separate exercise; this only gets you a repeatable way to prove it's happening in your own setup.

Leans on

Build one controlled test

Pick one fact and its question
Build three prompts, one seat moved

Run it and score it honestly

Ask the same question three times
Score each reply against the source

Confirm it, then keep it

If nothing failed, make it harder
Check your shape against published numbers
Save the three replies with the prompt

Where to go next

Paste only the specific part of a long document that answers the question

go there once you've confirmed the bias, to cut the document down instead of moving facts around inside it

Checked 15 Aug 2026

Part of the Large Language Models (LLMs) cookbook