How to Fix Claude Ignoring System Prompts & Rules

Why Claude Ignores System Prompts and Custom Instructions

When Claude ignores your system prompts, rules, or Project custom instructions, it is usually due to a phenomenon called attention dilution or recency bias.

As a conversation grows longer, or if you upload large files to a Project, the actual instructions you want Claude to follow can get lost in the noise of the active chat. Claude tends to prioritize the most recent instructions (at the very bottom of the prompt) or gets confused by conflicting information within uploaded documents. Fortunately, Claude is uniquely sensitive to structured prompt engineering, and reorganizing your input can quickly restore strict adherence to your rules.

Use these step-by-step methods to force Claude to respect your system prompts and custom instructions.

---

Step-by-Step Fixes to Enforce System Prompts

1. Structure Instructions Using XML Tags Claude is trained specifically to recognize and prioritize content wrapped in XML tags. Plain English paragraphs of instructions can blur together, but XML tags clearly define the boundaries of your rules.

Instead of writing: *Don't use jargon, write in tone X, and format as markdown.*

* Structure it like this: `xml <system_instructions> - Do not use industry jargon. - Maintain a highly professional, objective tone. - Format all output using clear markdown headers. </system_instructions> ` * Refer back to these tags in your main prompt: "Generate the report based strictly on the rules specified in <system_instructions>."

2. Reposition Critical Rules to the End (Recency Bias) If your prompt is long—especially if it contains source text, articles, or code blocks—Claude might lose track of instructions placed at the very beginning.

• Move your most critical constraints to the very bottom of the prompt, right before you submit it.
• Example Prompt Structure:
• *Context:* "Here is the background information: [Insert text]"
• *Task:* "Please summarize this background information."
• *Constraints (at the bottom):* "Remember: Do not exceed 150 words. Format the summary as a bulleted list. Do not use the word 'delve'."

3. Implement Few-Shot Examples If Claude is struggling to follow a complex system formatting rule, simply telling it what to do may not be enough. Show it exactly what a correct response looks like.

Provide one or two examples of a correct output inside XML tags in your system prompt:

`xml <example_input> Format this as a list. </example_input>

<example_output> * First key metric * Second key metric </example_output> ` This gives the model a clear structural template to match.

4. Remove Conflicting Documents or Old History If you are using Claude Projects, double-check your uploaded reference files. If an uploaded document contains old formatting guidelines or contradictory instructions, Claude may struggle to resolve the conflict and default to its standard behavior.

• Review the files in your Project's knowledge base and remove outdated reference material.
• If you are in a standard chat, start a fresh conversation. Once a chat thread becomes bloated, Claude's ability to stick to early system instructions drops significantly.

5. Explicitly Assign a Persona Claude responds exceptionally well to role-playing prompts. By establishing a professional context, you narrow the scope of the model's behavioral pathways.

• Instead of: "Write a code review."
• Use: "You are a strict, senior security engineer reviewing code for potential vulnerabilities. Your output must only contain high-priority security fixes. Do not comment on style or formatting."

---

When to Escalate

If Claude continues to ignore basic instructions across multiple brand-new chats, or if custom instructions in Projects are completely ignored despite using XML tags, the system may be experiencing a backend deployment issue or a quality regression.

• Submit Negative Feedback: Click the "Thumbs Down" button on the specific response where Claude failed to follow instructions. Provide details about which rule was ignored. This directly alerts Anthropic's engineering team to model alignment issues.
• Check Community Outages: Search developer forums or the official Anthropic Discord to see if other users are reporting sudden instruction-following regressions after a model update.