You are deep in a debugging session with Claude. The responses start feeling slightly off — the model seems to have forgotten something you established earlier. Then a message comes back that ignores a constraint you defined at the start of the conversation. Eventually Claude stops responding altogether or tells you the conversation is too long.
This is what a full context window looks like in practice. Understanding what is happening — and why the degradation starts before the hard limit — helps you work around it before it becomes a problem.
What the context window actually contains
Claude's context window is not just your messages and Claude's responses. It contains everything in the conversation in order:
- →Your system prompt or custom instructions if you have set them
- →Every user message in the conversation
- →Every Claude response in the conversation
- →Any files or documents you have pasted
- →Tool use outputs if you are using Claude's tool features
All of this accumulates as tokens. Claude Sonnet 4 has a 200,000 token context window. That sounds large until you paste a 500-line file (roughly 30,000 tokens), get several detailed responses (another 20,000-40,000 tokens), and realize you have consumed 25-35% of your context window in a single exchange.
Why degradation starts before the limit
The most important thing most developers do not know: Claude's quality degrades significantly before the context window is full.
Research on large language models consistently shows that attention — the mechanism models use to "remember" earlier content — weakens for content that is far back in the context. In a very long conversation, Claude can technically "see" everything, but its effective attention to content from early in the conversation diminishes as more content is added.
Practically, this means:
- →Constraints you defined at the start of the conversation get progressively ignored
- →Earlier architectural decisions get forgotten
- →The model may contradict something it agreed to 20 messages ago
- →Code quality tends to decrease as conversations run longer
This degradation typically becomes noticeable around 60-70% of the context window, well before the hard limit. By the time the context is 80% full, you are often getting meaningfully worse responses than you would get in a fresh conversation with a good summary.
The two limits you need to track
Claude has two rate limits that are separate from the context window:
The 5-hour session limit tracks your message volume within any rolling 5-hour window. This resets independently of the context window and resets on a rolling basis from your first message.
The 7-day weekly limit tracks cumulative usage across 7 days. Heavy users can hit this limit even when their context window is empty.
You can be rate limited without having a full context window, and you can have a full context window without being rate limited. Both need to be monitored independently.
TokenPulse shows all three — context window percentage, 5-hour session utilization, and 7-day utilization — simultaneously in the popup, which is the only way to see the full picture at once without guessing.
The right time to start a new conversation
The common mistake is waiting until Claude tells you the conversation is too long. By that point you have already been getting degraded responses for a while.
The right trigger is 60% of the context window. At 60%, you still have enough headroom for Claude to give you a useful summary of the conversation. Ask for it, start a new conversation, paste the summary, and continue from there.
The summary should include:
- →The problem you were solving and any constraints established
- →Key decisions already made and the reasoning behind them
- →Current state — what has been implemented, what is still to do
- →Any specific code snippets or data structures that need to carry forward
This summary approach typically condenses 40,000 tokens of conversation into 1,000-2,000 tokens while preserving everything that matters for continuity.
How to see your context window in real time
Claude does not show you the context window percentage in the interface. The only way to see it is through the internal API that Claude uses to track usage — which is what TokenPulse reads.
The live bar above the input box updates after every message, showing you the percentage in real time. When it hits 60%, that is your signal to summarize and restart. When it hits 75%, the built-in notification fires so you cannot miss it even if you are focused on coding.
Summary
- →Claude's context window contains all messages, responses, and pasted content — it fills faster than most developers expect
- →Quality degrades noticeably at 60-70% capacity, not just at the hard limit
- →Claude has two separate rate limits (5-hour and 7-day) that are independent of the context window
- →The right time to restart a conversation is at 60% — not when Claude tells you it is too long
- →TokenPulse shows context window, session limit, and weekly limit simultaneously in real time