how to stop claude ignoring instructions

It can be incredibly frustrating when Claude bypasses specific constraints you have set, such as word counts, formatting rules, or stylistic prohibitions. This behaviour, often referred to as instruction drift, typically happens when a prompt lacks clear structural hierarchy or when the conversation window becomes too cluttered for the model to prioritising your initial rules.

Because Claude is trained differently to other large language models, it requires a specific prompting structure to remain highly compliant. Here is how to stop Claude from ignoring your instructions and get it back on track.

Use XML tags to establish structure

Claude is explicitly trained to recognise and respect structure within XML tags. If you dump instructions, context, and examples into a single block of text, the model can struggle to differentiate between the data it is supposed to process and the rules it must follow.

Wrap your core instructions in tags like <instructions></instructions> or <rules></rules>. For example:

<instructions>
- Write in a professional, neutral tone.
- Do not use passive voice.
- Limit the output to exactly three paragraphs.
</instructions>

<text_to_process>
[Your input text here]
</text_to_process>

This clear separation prevents Claude from confusing your guidelines with the raw data it is analyzing.

Exploit recency bias by moving instructions to the end

Like many language models, Claude exhibits a level of recency bias, meaning it pays closer attention to the text at the very end of a prompt. If you place a long set of rules at the beginning, followed by thousands of words of reference material, Claude is highly likely to ignore your initial constraints.

Always place your most critical instructions, formatting constraints, and final execution commands at the very bottom of your input. This ensures those rules are the last thing the model reads before it begins generating its response.

Start a clean conversation to clear the context window

As a chat session grows longer, Claude has to process an increasingly massive amount of historical text. If the model previously ignored a rule in that chat, or if the conversation has drifted through several topics, Claude's attention mechanism becomes diluted, causing it to lose track of your original prompt constraints.

If Claude begins ignoring instructions mid-conversation, copy your core prompt, start a completely fresh chat session, and paste it there. If you find the interface laggy or experience issues where you cannot load a new chat, see our guide on Claude not responding to troubleshoot browser and session state issues.

Use explicit negative constraints

Models often struggle to interpret vague, positive instructions. Instead of asking Claude to "keep it simple" or "write normally," use highly specific, negative constraints that tell the model exactly what to avoid.

Explicitly list banned words, forbidden phrases, or prohibited formatting styles. For instance, instead of "write in a friendly tone," try "Do not use corporate jargon, do not use exclamation marks, and do not use the word 'delve'." Giving the model a clear list of what *not* to do is often far more effective than general stylistic requests.

Separate system instructions from user inputs in the API

If you are using Claude via the Anthropic API or a developer console rather than the standard web interface, ensure you are utilizing the dedicated system parameter correctly. Developers often mistakenly bundle system instructions into the user message array.

Putting your core rules into the system parameter ensures they are weighted heavily across the entire API call lifecycle. If your API calls are failing to return any output or cutting off entirely while you test these configurations, check our walkthrough on resolving the Anthropic API 529 error to rule out server capacity bottlenecks.