How to Stop Claude From Forgetting Instructions

It is frustrating when Claude starts a conversation adhering strictly to your formatting rules, style guides, or constraints, only to drift and begin ignoring them 10 or 15 prompts later. This behavior occurs because of "attention degradation" over long context windows. As your chat history grows, earlier messages receive less weight from the model's attention mechanism, causing it to prioritize recent turns over your original instructions.

To keep Claude aligned throughout a long workspace session, follow these practical troubleshooting steps.

1. Structure instructions with XML tags Claude is trained specifically to recognize and prioritize structured data wrapped in XML tags. Plain-text rules like "Do not use emojis and keep your answers brief" are easily lost in long contexts. Instead, format your initial prompt using clear XML containers:

`xml <system_instructions> <rule>Do not use emojis under any circumstances.</rule> <rule>Limit all responses to under 150 words.</rule> <rule>Always format output using Markdown lists.</rule> </system_instructions> `

This structured format makes it easier for Claude's attention mechanism to isolate and reference the rules, even as the chat history grows.

2. Utilize Claude Projects to pin instructions If you are a Claude Pro or Team subscriber, do not rely solely on your starting prompt to carry rules throughout a long thread. Instead, use the **Projects** feature: 1. Open your Claude dashboard and create a new **Project**. 2. In the right-hand panel, click **Set custom instructions**. 3. Paste your system-level rules and formatting constraints here. 4. Click **Save**.

Project instructions are injected directly into the system prompt of every single message you send within that project. This ensures your rules never get pushed out of Claude's active context window, no matter how long the thread becomes.

3. Implement periodic prompt reinforcement When you notice Claude beginning to drift or forget rules, do not just correct the immediate mistake. You must explicitly refresh its memory. Use a technique called "prompt reinforcement" on your next turn:

• Incorrect correction: "You used an emoji. Fix that."
• Correct reinforcement: "Correction: You violated the formatting rule <rule>Do not use emojis</rule>. Rewrite your last response without emojis, and keep this rule active for all subsequent turns."

Providing the context of the rule alongside the correction forces the model to realign its attention parameters to the original instruction set.

4. Run regular context pruning (New Chats) Every message, response, and error in your chat history consumes tokens and degrades Claude's focus. If you have been troubleshooting or brainstorming in a single thread for hours, the context window is bloated.

When Claude starts repeatedly failing to follow instructions, it is time to prune: 1. Copy the current working state of your task (e.g., the latest draft, code block, or outline). 2. Click Start New Chat (or use the shortcut Ctrl/Cmd + K). 3. Start the new chat by pasting your original XML rules followed by the copied working state: *"Here is our current progress: [paste state]. Continue developing this using the following rules: [paste XML rules]."*

This resets Claude's attention window completely, removing historical noise while preserving your progress.

5. Pass instructions in the API system parameter If you are accessing Claude via the Anthropic API rather than the web interface, do not pass your behavioral rules in the `user` message role.

Always place long-term instructions, personas, and strict constraints inside the system parameter of your API request. The API architecture weights the system parameter differently, ensuring that these instructions maintain high priority even when the conversation history approaches the context limit.

When to escalate If Claude begins ignoring your instructions within the first two or three prompts of a brand-new, clean thread, or if it outright refuses to follow system prompt rules that previously worked, there may be an active platform regression or a bug in Anthropic's prompt-parsing pipeline. Check the official Anthropic Status page to see if there is an active incident. If services are operational, use the "thumbs down" feedback button on the web app to submit the specific conversation to Anthropic for evaluation.