Fix Claude Forgetting Instructions in Long Chats

As conversation history grows, large language models like Claude can experience "attention drift." Even though Claude has a massive context window (over 200,000 tokens), its ability to recall and strictly adhere to instructions given at the very beginning of a long chat naturally degrades. This is known as the "loss in the middle" phenomenon, where the model prioritizes the most recent messages over earlier system instructions.

When this happens, Claude may stop formatting output correctly, drop requested constraints, or slip back into its default writing style instead of your specified persona.

Why Claude loses track of early instructions

There are three main technical reasons why Claude begins to ignore your rules as a thread grows longer: 1. Recency Bias: The model assigns higher mathematical weight to the tokens closest to its current generation point (the last few prompt turns). 2. Context Dilution: If your chat contains hundreds of lines of code or raw data, your original formatting instructions get buried under thousands of words of distracting context. 3. Implicit Reinforcement: If Claude makes a minor formatting error and you do not correct it immediately, its own incorrect output becomes part of its historical context, reinforcing the wrong behavior in subsequent replies.

How to fix instruction drift in Claude

To restore instruction adherence without losing your chat history, work through these steps:

1. Re-anchor your instructions using XML tags Claude is trained specifically to recognize and prioritize instructions wrapped in XML tags. If Claude begins to drift, send a structural correction message using clear tags to reset its focus.

Copy and paste this template into your chat: `xml <system_reminder> You are ignoring the original formatting and behavior rules established at the beginning of this chat. Please review and apply these rules strictly to all future responses: - [Insert rule 1, e.g., "Do not write conversational filler"] - [Insert rule 2, e.g., "Output code in Python only"] </system_reminder> ` Wrapping rules in explicit XML tags signals to Claude's attention mechanism that these inputs are metadata constraints, not standard conversational text.

2. Set up a Project with Custom Instructions (Claude Pro/Team) If you are using Claude Pro or Claude Team, do not rely solely on the active chat box for long-term instructions. Use the **Projects** feature: 1. Click on **Projects** in the left sidebar and create a new project. 2. In the right-hand panel, click **Set custom instructions**. 3. Paste your core rules, formatting requirements, or persona constraints here.

Custom instructions in Projects are injected as a system prompt at the beginning of *every single message exchange*, forcing Claude to evaluate them with high priority regardless of how long the chat becomes.

3. Clear the historical context by starting a new thread Once a chat thread reaches a certain length, the cumulative tokens degrade response quality and increase latency. If Claude continues to fail despite reminders, you must migrate to a clean slate: 1. Ask Claude to summarize the current state of your work: *"Summarize our progress, decisions, and current code version so far so I can paste it into a new chat."* 2. Copy the summary. 3. Click **New Chat** (or `Ctrl/Cmd + K`). 4. Paste the summary along with your original formatting instructions in the very first prompt.

4. Correct errors immediately Never let Claude get away with a bad response. If Claude ignores a rule, immediately reply with: *"You broke rule X. Redo your last response and fix this."* If you continue the conversation without correcting the mistake, Claude treats its own flawed output as the new standard for the thread.

When to escalate

If Claude completely ignores simple, direct instructions even in brand-new, short chat threads, this points to a system-wide model degradation or an update regression. Check the Anthropic status page to see if a model update or partial outage is currently affecting generation quality. If the system is fully operational but Claude continues to ignore basic prompts across all new sessions, contact Anthropic Support via the in-app help bubble to report a model-behavior regression.