This ‘anti-goal’ immediate trick retains ChatGPT from going rogue

All too usually, ChatGPT, Claude, and Gemini overstep their directions as a result of they’re so centered on making you cheerful. For instance, an AI might leap forward and fully rewrite a doc when all you needed was some centered suggestions, or it could draft a brand-new recipe once you simply needed assist substituting an ingredient. 

You would possibly assume the answer is to inform the AI chatbot what it shouldn’t do in your immediate. However that would backfire. In its rush to supply a useful response to the primary a number of phrases in your request, it could deemphasize your restrictions—or skip them fully.

The trick is to construction your immediate in a manner that forces the AI to cease and provides every a part of your request its due weight. You possibly can even tag every phase of your immediate to inform the AI what their functions are.

AI chatbots like ChatGPT, Claude, and Gemini are able to recognizing “markup” parts that present hints to the construction of your request. With markup parts, you’re primarily saying issues like “This primary a part of my immediate is background context for my request” and “This center a part of my immediate comprises directions on what to do” and “This final a part of my immediate is telling you what not to do.”

We’re going to make use of XML markup, and we’re going to make use of open/shut units of XML tags to outline the areas of a immediate. On this instance, one set will outline the “objective” of the immediate (e.g., what you need the AI to do) and one other set will outline the “anti-goal” of the immediate (e.g., what you don’t need the AI to do).

Let’s say you need assist writing a canopy letter for a job software however you don’t need the AI to jot down (or rewrite) the letter for you. As a substitute, you merely need suggestions on the letter you’ve already written.

Right here’s an instance “anti-goal” immediate that ought to maintain the AI in line:

Professional Profession Coach and Hiring Supervisor

Assessment the offered cowl letter for the [Insert Job Title] place. Consider whether or not my introduction is participating, if my tone is suitable for the business, and if I successfully join my previous expertise to the core necessities of the function. Please level out any sections the place my claims sound too generic or lack particular proof.

Do NOT rewrite any a part of the duvet letter. Do NOT present a revised or “improved” draft. Do NOT deal with minor typos or grammar. Present your entire suggestions as actionable critiques in a bulleted listing.

This immediate, which additionally consists of XML tags that inform the AI what its “function” ought to be for this immediate, ought to maintain the AI centered on its main activity (providing you with suggestions in your cowl letter) whereas stopping it from doing undesirable work (leaping forward and re-writing your letter).

Right here’s one other instance:

Journey Information
 

Recommend 5 distinct neighborhoods to remain in whereas visiting Tokyo for every week. Give attention to areas with nice avenue meals and straightforward transit entry. 
 

 
Do NOT construct a day-by-day itinerary. Do NOT recommend particular accommodations or eating places. Hold the outline of every neighborhood beneath three sentences.

I attempted this immediate with the ChatGPT’s new GPT-5.5 Instantaneous mannequin and obtained this clear, centered response:

Shinjuku
A high-energy hub with limitless late-night meals alleys, standing bars, and yokocho streets full of informal bites. Transit is unbeatable because of Shinjuku Station, making it straightforward to achieve practically each a part of Tokyo.

Ueno
Recognized for its conventional environment, open-air markets, and reasonably priced avenue snacks round Ameya-Yokocho. Ueno Station connects a number of JR and subway strains, plus direct entry to Narita Airport routes.

Asakusa…

The takeaway right here is that telling your AI what not to do is among the strongest alerts to incorporate in a immediate, and doing it by way of structured XML tags helps make sure the mannequin will get the message.