Separate instructions from data with XML tags

A stray line inside a pasted email got followed as an instruction instead of yours.

BeginnerRead 5 minDoing: half an hour6 steps

▶ Start

What you'll have at the end

A prompt split into a labeled instructions section and a labeled data section that a model reliably tells apart

Before you start

You need

A prompt that already sends pasted or retrieved text, such as an email or a support message, straight into a model call with no separation from your own request.

Not covered

Defending against a determined attacker trying to hijack your prompt, with output filters, privilege limits, and red-teaming; this only stops the ordinary case where a harmless-looking line gets read as a command.

Leans on

See the mix-up for yourself

Reproduce the mix-up with your own prompt

Wrap and label the two halves

Split the prompt into two tagged blocks
State what the data block actually is

Stress-test the boundary

Plant a louder trap line and rerun
Give each pasted item its own tag

Make it stick in your code

Wrap every prompt the same way in code

Checked 15 Aug 2026

Part of the Context Engineering cookbook