You ask your AI for simple feedback—and suddenly it rewrites your entire document. You just want a quick substitution tip, and it hands you a brand-new recipe from scratch. Why does this keep happening?
The problem isn't that the AI is dumb. It's trying way too hard to please you. Eager to deliver a satisfying answer, ChatGPT, Claude, and Gemini often leap ahead, ignoring your actual constraints. And if you think telling them what *not* to do will fix this, think again—your restrictions often get buried under the flood of positive intent.
The real trick? Force the AI to stop and weigh every part of your request equally. You can do that by tagging each segment of your prompt, explicitly labeling its purpose. Think of it like leaving sticky notes for a brilliant but overeager assistant.
Modern AI chatbots understand "markup" elements—little structural hints that say, "This part is background context," "This is the main instruction," and "This is the hard no." By using open and close XML tags, you create a clear boundary between what you want and what you absolutely don't.
Let's call the positive goal your "goal" tag, and the forbidden zone your "anti-goal" tag. When you define both, the model's attention splits evenly, preventing it from racing ahead into unwanted territory.
Imagine you need feedback on a cover letter—but you absolutely don't want the AI to write or rewrite a single word. Here's how an anti-goal prompt keeps it in check:
Expert Career Coach and Hiring Manager
Review the provided cover letter for the [Insert Job Title] position. Evaluate whether my introduction is engaging, if my tone fits the industry, and if I connect my past experience to core role requirements. Point out any sections where my claims feel generic or lack evidence.
Do NOT rewrite any part of the cover letter. Do NOT provide a revised or improved draft. Do NOT focus on minor typos or grammar. Provide all feedback as a bulleted list of actionable critiques.
With XML tags wrapping the role, goal, and anti-goal, the AI stays laser-focused on critique—not creation.
Another example: You're planning a trip to Tokyo and need neighborhood suggestions—nothing more. Try this:
Travel Guide
Suggest 5 distinct neighborhoods to stay in while visiting Tokyo for a week. Focus on areas with great street food and easy transit access.
Do NOT build a day-by-day itinerary. Do NOT suggest specific hotels or restaurants. Keep each neighborhood description under three sentences.
When I tested this on ChatGPT's new GPT-5.5 Instant model, the response was clean and disciplined. Here's a snippet: Shinjuku—A high-energy hub with endless late-night food alleys and yokocho streets packed with casual bites. Transit is unbeatable via Shinjuku Station. Ueno—Known for traditional atmosphere, open-air markets, and affordable street snacks around Ameya-Yokocho. Ueno Station connects multiple lines and offers direct Narita Airport routes. Asakusa…
The lesson is simple: telling your AI what not to do is one of the most powerful signals you can give. And wrapping those instructions in structured XML tags ensures the message lands—loud and clear—so your assistant stays on task, not overeager.
