Why is Claude giving worse answers? How to fix quality issues

If you notice that Claude (especially Claude 3.5 Sonnet or Opus) has suddenly started giving generic, inaccurate, or lazy answers, you are likely experiencing the effects of context drift, prompt exhaustion, or temporary system optimization updates.

While AI models do not "get dumber" overnight, their behavior can change due to behind-the-scenes system prompt adjustments by Anthropic, or because a long-running chat session has become cluttered with contradictory information. This guide explains why Claude's output quality drops and how to restore its reasoning capabilities.

Why Claude's Output Quality Degrades

There are three primary reasons Claude's performance can decline during a session or across days:

• Context Window Pollution: In long chat threads, Claude remembers everything previously said. If the thread contains formatting errors, off-topic details, or incorrect AI attempts, Claude will copy those bad patterns, leading to a downward spiral in output quality.
• In-Flight System Prompt Adjustments: Anthropic occasionally updates the invisible "system instructions" that guide Claude's safety protocols and default formatting. These changes can temporarily make the model more verbose, overly cautious, or less direct.
• Prompt Fatigue: If your prompt lacks explicit constraints, Claude defaults to conversational, generalized answers rather than precise, technical ones.

How to Fix Poor Quality Answers

Follow these steps to reset Claude's behavior and force the model to generate high-quality, accurate responses.

1. Start a Fresh Chat Session Do not keep using a single chat thread for days. If Claude starts giving poor answers, click **New Chat**. This completely clears the active context window, forcing the model to evaluate your instructions without being influenced by past mistakes in the chat history.

2. Implement XML Tags for Structure Claude is natively trained to recognize and prioritize information enclosed in XML tags. If you are feeding the model data, code, or complex instructions, organize your prompt like this:

`xml <instructions> Write a python script that cleans the attached CSV file. Do not use external libraries. </instructions>

<source_data> [Paste data here] </source_data> ` This prevents the model from getting confused about what is an instruction versus what is raw data.

3. Use "Few-Shot" Prompting If Claude's style or logic has degraded, provide 1 or 2 examples of the exact output format you expect. Showing Claude what a "good" answer looks like is the single most effective way to correct a quality regression.

4. Direct the Thinking Process (Chain of Thought) For complex reasoning or coding tasks, explicitly tell Claude to think before answering. You can prompt the model to use a specific thinking space:

• *"Please think step-by-step inside <thinking> tags to analyze the problem before providing your final answer in <response> tags."*

This forces the model to allocate computational focus to the logic before drafting the final text.

5. Strip Conversational Filler If Claude is giving overly polite or shallow answers, explicitly suppress this behavior in your prompt. Add this directive to the end of your message:

• *"Be concise. Avoid conversational filler, introductions, and apologies. Provide only the requested technical output."*

---

When to Escalate

If you have started a fresh chat, used structured XML prompts, and provided explicit examples, but Claude still returns completely broken code, hallucinated facts, or repetitive refusal loops:

• Check the Anthropic Status Page: Visit status.anthropic.com to see if there is an active performance degradation incident.
• Report the Bug: Click the Thumbs Down icon directly beneath the bad response in Claude.ai. This sends the specific context directly to Anthropic's engineering team for model evaluation.
• Clear Browser Storage: If the UI itself feels slow or buggy, clear your browser cache and cookies, then log back in.