Claude Forgetting Instructions? How to Stop Prompt Drift

As conversations with Claude grow longer, you may notice the model starts ignoring earlier constraints, formatting guidelines, or behavioral rules. This issue—often called "prompt drift" or "attention decay"—occurs when the sheer volume of chat history dilutes the structural weight of your initial instructions.

Because Claude prioritizes more recent messages in its context window, older instructions lose their influence over time. Follow these steps to anchor your instructions and ensure Claude maintains consistent behavior throughout your session.

1. Structure instructions with XML tags Claude is trained specifically to recognize and prioritize content structured inside XML tags. Instead of writing instructions as plain text paragraphs, wrap your constraints, rules, and rulesets inside explicit tags. This signals to Claude's attention mechanism that these blocks are structural rules rather than conversational filler.

Format your prompts like this:

`xml <system_instructions> - Output only in valid JSON format. - Do not include conversational filler or introductions. - Write in a professional, technical tone. </system_instructions>

<user_query> [Your actual request here] </user_query> `

Using tags like <instructions>, <rules>, or <constraints> helps the model separate the "how" from the "what" even as the chat history grows.

2. Re-prime the model at the end of the prompt If you are sending a long prompt with a lot of reference data, place your operational instructions at the very end of the message. If you put instructions at the beginning, then paste 1,500 words of raw data, Claude is more likely to lose track of the instructions.

Always structure your input sequentially: 1. Context/Reference Data 2. User Query 3. Final constraints and instructions (the "reminders")

Example: "...[Paste reference text here]...

Based on the text above, write a summary. Remember to adhere to the formatting rules in the <system_instructions> tag at the top of this chat."

3. Leverage Claude Projects and Custom Instructions If you are using Claude Pro or Team, do not rely on standard chats for long-running workflows. Use the **Projects** feature instead.

1. Navigate to the Claude sidebar and click Projects.
2. Create a new project and open the Project Instructions panel on the right.
3. Input your behavioral rules, formatting preferences, and system guidelines here.

Instructions placed in the Project Instructions are appended to *every single message* behind the scenes. This prevents prompt drift because the rules are injected alongside each new input, bypassing the attention decay that occurs in normal chats.

4. Use the API's System Parameter If you are accessing Claude via the Anthropic API or the Console Workbench, do not pass your rules inside the `user` message role. Instead, use the dedicated `system` parameter.

`json { "model": "claude-3-5-sonnet-20241022", "max_tokens": 1024, "system": "You are a strict code validator. Only output executable Python code blocks inside markdown.", "messages": [ {"role": "user", "content": "Write a function to parse CSV files."} ] } `

The system parameter acts as a global anchor, keeping instructions active across all subsequent turns in the API call.

5. Implement context clearing and summarizing When a single chat thread reaches 15 or more interactions, Claude's internal attention is split across too many tokens. To fix this:

1. Ask Claude to summarize the work completed so far: *"Provide a concise summary of our progress, decisions made, and the current state of our code/text."*
2. Copy that summary.
3. Start a brand-new chat session.
4. Paste the summary and re-insert your original XML-tagged instructions. This resets the attention span while preserving your progress.

When to escalate If Claude continues to ignore explicit, simple instructions even in a brand-new chat session using XML tags, there may be an active quality regression issue. Check the official Anthropic Status page (status.anthropic.com) to see if there is an active incident involving model performance or degraded text generation quality.