Fix Claude Ignoring Negative Instructions ("Do Not" Rules)

Understanding Negation Bias in Claude

One of the most common user frustrations with LLMs like Claude is their tendency to ignore negative instructions. You might write *"Do not include any introductory remarks"* or *"Do not use the word 'delve' in this article,"* only for Claude to start its response with *"Certainly, let's delve into..."*

This happens because of how transformer models process tokens. When Claude parses your prompt, words like "delve" or "introductory" are activated in its attention mechanism. The negation modifier ("do not", "never", "exclude") has a weaker semantic weight than the core noun or verb itself. Essentially, telling Claude *not* to think of a pink elephant makes the concept of a pink elephant highly active in its immediate context.

To successfully force Claude to adhere to negative constraints, you must change how you structure your instructions. Follow these steps to resolve the issue.

---

Steps to Force Claude to Follow Negative Rules

1. Reframe Negatives as Positive Actions Instead of telling Claude what *not* to do, tell it exactly what it *should* do instead. This keeps its attention focused on the correct output pattern. * **Bad negative prompt:** *"Do not write a long introduction."* * **Good positive prompt:** *"Begin the response immediately with the first paragraph of the main body. Start with the word 'The' and directly address the core topic."* * **Bad negative prompt:** *"Do not use passive voice."* * **Good positive prompt:** *"Write every sentence using active voice, where the subject clearly performs the action."*

2. Isolate Constraints Using XML Tags Claude is specifically trained to recognize and prioritize structural rules wrapped in XML tags. By placing your forbidden items inside clear XML structures, you elevate their importance in Claude's attention map.

Example format: `xml <strict_constraints> - Forbidden words: "delve", "testament", "moreover" - Forbidden formats: No bullet points, no introductory pleasantries, no markdown headers. </strict_constraints>

<task> Write a 200-word product description. Before outputting, check your draft against the <strict_constraints> block. If any forbidden items are present, rewrite the sentence to eliminate them. </task> `

3. Implement Few-Shot Examples (Negative Training) Show Claude examples of what a bad output looks like versus what a corrected, compliant output looks like. This is highly effective for complex formatting or stylistic constraints.

Example: `text Here is an example of an incorrect response containing forbidden words: "Let us delve into the features of this tool."

Here is the correct version of that response (with forbidden words removed): "This tool offers several key features."

Now, generate my request following this pattern: [Insert prompt] `

4. Build a Self-Correction Step Into the Prompt Instruct Claude to analyze its own draft before presenting the final response. This forces the model to run a secondary "compliance check" within its generation window. * Add this block to your prompt: *"Before outputting your final answer, write a brief, hidden draft. Review the draft against my negative constraints. If you find any violations, edit them out and output only the final, cleaned version."*

5. Use the System Prompt or Projects Panel If you are using Claude Pro or Team, you can set account-wide or project-wide rules. Placing negative constraints in the "System Prompt" or "Project Instructions" panel gives them global authority over every turn of the conversation, making them much harder for the model to forget.

---

When to Escalate

If Claude continues to ignore highly explicit, simple constraints despite using XML tags and positive framing, there may be an API configuration error or a temporary regression in the underlying model version you are accessing. Check the Anthropic Status page (status.anthropic.com) to verify that there are no active API service interruptions or degraded performance flags. If you are using third-party integrations (like Poe, TypingMind, or custom API wrappers), ensure their system prompts are not overriding your negative instructions. If the problem persists on the official Claude.ai interface, contact Anthropic Support through the dashboard help icon.