Replace negative prompting with instructions a model will follow

I told the chatbot never to mention a competitor, and it worked the name into a reply anyway.

Beginner to intermediateRead 5 minDoing: about 35 minutes7 steps

What you'll have at the end

A rewritten instruction, retested across five fresh conversations, that stayed clean where the original violated three times out of five, with every other 'don't' line in the same prompt converted the same way.

Before you start

You need

A chatbot or LLM app whose system prompt you can edit, with at least one 'do not' instruction you can already watch the model break.

Not covered

Building an automated test harness that reruns this check on every future prompt edit, instead of the manual five-trial spot check here.

Leans on

Prove the fix actually works

Reproduce the failure across five fresh tries
Name the one thing to say
Retest the rewrite the same way
Line up before and after, check the rest

Apply the fix everywhere it hides

Scan the rest of the prompt
Give the one hard boundary a script

Make it a standing habit

Write down the standing rule

Where to go next

Audit a failing prompt against a checklist of common mistakes

If the positive rewrite still gets violated, run the whole prompt through a broader mistake checklist before assuming this technique failed.

The same prompt can work on one chatbot and fail on another

If the same rewritten instruction holds on one chatbot and fails on another, that's a model difference to check separately, not a flaw in the rewrite.

Checked 18 Aug 2026

Part of the Prompt Engineering cookbook