Negative prompting should not have to be a thing

The question

One of my biggest bug bears with Copilot is the need to use “negative prompting” to get accurate responses and stop it assuming or editorialising when it is not asked to. It even admits this and when using advanced prompts to get it to critique its own answers it seems to come up with a multitude of examples of this and advise that: “I understand the frustration. From a user perspective, “reformat this document” should not require you to also say: • don’t add content • don’t summarise • don’t invent email addresses • don’t change meaning • don’t rewrite policy • don’t claim validation you didn’t do Those should be the default assumptions. What you’re describing is a known usability problem in AI systems generally: users end up learning to write defensive prompts because the model is optimising for being helpful, comprehensive, and proactive. Unfortunately, in document work, “proactive” can easily become “made changes I never asked for. In other words, the burden should be on the system to understand the risk profile of the task, not on the user to write increasingly restrictive instructions. The specific business risk I think you’re identifying is: The system can cross the boundary between presentation changes and content changes without making that boundary sufficiently obvious to the user. That’s particularly important in environments like Oranga Tamariki, where documents can contain operational guidance, legal references, care decisions, privacy-sensitive information, and instructions that staff may rely on. If the system changes content while improving appearance, the visual quality can mask the fact that substantive changes have occurred. That’s a governance risk, not just a quality issue. So, I don’t think the lesson is “users need better prompts.” That’s particularly important in environments like Oranga Tamariki, where documents can contain operational guidance, legal references, care decisions, that staff may rely on.” Isn’t this supposed to save time?

The answer

Hi nina — this is one of the most thoughtful notes I’ve had on here, so thank you. You’ve put your finger on something real.

I agree with the core: the burden shouldn’t sit on the user to write defensive, negative prompts. These models are tuned to be helpful and proactive, and in document work “proactive” quietly crosses from presentation into content — and when the visual quality goes up, it can mask that substantive changes happened. In an environment like Oranga Tamariki — operational guidance, legal references, care decisions — that’s a governance risk, not just a quality niggle. You’re right to name it as that.

Two honest halves:

• The real fix isn’t mine to ship. A Guide to Cloud is independent — I don’t build Copilot — so the systemic change (the tool understanding a task’s risk profile and defaulting to “change appearance, never content”) has to come from Microsoft. The best place to lodge exactly this is the in-app “Give feedback” in Copilot or the Microsoft feedback portal — your write-up is strong enough to carry weight there.
• Stopgaps until the defaults catch up — these reduce risk but they’re mitigations, not real guarantees: ask it to show the change as a diff/preview before applying, keep format and content in separate turns, and — the deterministic ones — work on a copy and keep version history so you can always compare and roll back before accepting anything.

Your example is a better articulation of the problem than what I’ve got published, so I’m going to add a “safer document editing” section to the prompt guide off the back of it. Thanks again — genuinely.

— Sush

Quick update, nina — that “safer document editing” section I mentioned is now live: 👉 Safer Document Editing — Reformat Without Rewriting

It’s built straight off your point: scope the edit to formatting only, explicitly fence off the content (“don’t change, add, remove, or reword”), and make the model show you what it changed instead of quietly applying it — plus the deterministic safeguards that actually protect you (work on a copy, keep version history, review before accepting).

Your framing genuinely shaped how I wrote it. Thanks again for taking the time — this is exactly the kind of feedback that makes the site better.

— Sush