Fix Claude Getting Dumber or Unresponsive in Long Chats

As you continue a single conversation with Claude over many hours or days, you might notice a sudden drop in response quality. The model might start ignoring your formatting rules, repeating previously used phrases, hallucinating incorrect facts, or failing to follow simple negative constraints (e.g., "do not write explanations").

This behavior is a known limitation of large language models called "context drift" or "attention dilution." This guide explains why this quality regression happens and how to restore Claude's performance.

Why Claude's Performance Degrades in Long Conversations

To understand why Claude "gets dumber" over time, you need to understand how its context window works:

• The Attention Tax: Every time you send a new message in a chat thread, Claude does not just read your latest message. It re-reads the *entire* conversation from the very beginning.
• Information Overload: As the thread grows, the model has to process hundreds of thousands of words of history. Important instructions you gave in your very first prompt get buried under pages of subsequent dialogue, code revisions, and system outputs. The model's "attention" is stretched thin, causing it to prioritize recent chat history over early core instructions.
• Reinforcing Mistakes: If Claude made a small formatting error five messages ago and you didn't correct it, that mistake remains in the context history. Claude sees its own past mistake as part of the established pattern for the conversation, making it highly likely to repeat the error.

How to Restore Claude's Answer Quality

1. Start a Clean Chat Session with a Summary Starting a fresh chat is the single most effective way to restore Claude's intelligence. However, you don't have to lose your progress. Use a "bridge prompt" to migrate your work.

1. In your long, degraded chat, ask Claude: "Summarize everything we have completed in this thread so far. Highlight the core decisions made, the current state of our code/text, and the rules we are following. Format this as a concise system prompt for a new chat."
2. Copy Claude's summary.
3. Click Start New Chat.
4. Paste the summary as your first message in the new chat, followed by your next instruction.

2. Edit Previous Messages Instead of Replying To prevent a chat from growing too long in the first place, use the "Edit" feature on your own prompts instead of adding new messages when correcting code or text.

1. Hover over your last prompt and click the edit icon (pencil symbol).
2. Change your prompt (e.g., adjust the requirements) and submit.
3. This action completely deletes the branch of the conversation that went off-track, keeping the total active context history short and focused.

3. Clear Outdated File Attachments If you uploaded PDFs, CSVs, or text documents early in the conversation, those files are still being re-processed with every single message you send.

• If you no longer need to reference those specific documents, start a new chat without them.
• If you only need small snippets of those files, copy and paste just the relevant sections rather than re-uploading entire multi-megabyte documents.

4. Use XML Tags to Refocus Attention If you must use a long thread, help Claude focus its attention by wrapping your current instructions in clear structural containers like XML tags. This helps the model separate your immediate instruction from the historical noise.

* Example: ` <current_instruction> Please rewrite the last function, but ensure you do not use deprecated methods. Follow the rules defined at the start of our chat. </current_instruction> `

When to Escalate

If you experience severe quality regressions—such as gibberish outputs, extreme repetitiveness, or refusal to answer simple questions—even in brand-new chat sessions, the issue is not context drift. This points to a model deployment error or a temporary regression in Anthropic's production servers. Check developer forums or the official Anthropic status page to see if a system-wide model degradation is actively being resolved. If the issue persists across different devices and new chats, contact Anthropic support through the help widget in your Claude console.