Introduction
Here is what most people do not realise. Claude does not just read your latest message. Every time you send a message, Claude re-reads the entire conversation from the very beginning before responding. Message 1 costs almost nothing. By message 20, Claude is re-reading 19 previous exchanges before it even starts on your new question. But this can be fixed.
Here are 10 practical changes that reduce how quickly you burn through your limit without reducing the quality of what you get from Claude.
Keep Conversations Short and Focused
Start a fresh chat every 15 to 20 messages
Long conversations are token furnaces. When a session gets heavy, ask Claude to summarise the key decisions and next steps, copy that summary, open a new session, and paste it as your first message. You carry the context forward without Claude re-reading 30 exchanges every time you type a new question.
Start a new chat when the topic changes
If you asked Claude to help with a presentation, then a data model, and then an email in the same chat, Claude is still re-reading the presentation conversation every time it thinks about your email. Old unrelated context is dead weight, burning tokens on every response. New topic means new chat, always.
Fix How You Correct Claude
Edit your prompt instead of sending a follow-up
Every time you type "No, I meant..." or "Actually, change X to Y," you add another message to the conversation, and Claude re-reads everything again. In Claude Chat, there is a pencil icon on every message. Click it, rewrite your original prompt, and regenerate. The old exchange gets replaced, not stacked on top. This single habit makes a noticeable difference within the first day.
Never ask Claude to redo the whole thing
When one section of a report is wrong, do not say "redo the report." Say "only redo section 3, keep everything else." Every full redo regenerates the entire output and burns those tokens again. Point to the specific section and tell Claude exactly what is wrong. Adding "no commentary, just the output" when you know what you want also helps, since every line of "Happy to help! Here is what I did..." is a token you are paying for.
Use the Right Tool for the Right Task
Match the model to the task
Sonnet handles grammar checks, brainstorming, summarising, and short drafts at a fraction of the cost of Opus. Opus with Extended Thinking is for deep, complex, multi-step work. A useful rule of thumb is that if Claude answers in under 30 seconds, it probably did not need Opus. Switch models before you start the session, not after you have already burned through your allocation.
Turn off features you are not actively using
Web search, connectors, and Extended Thinking consume tokens even when you do not actively need them for the task in front of you. Writing your own content? Turn off search. Doing a grammar check? Turn off Extended Thinking. The default should be everything off. Turn features on per task, not per session.
Batch your tasks into one message
Three separate prompts equal three full context reloads. One message with three tasks equals one reload. Instead of sending "summarise this," then "list the key points," then "suggest a title" as separate messages, combine them into one. The results are often better because Claude sees the full picture at once.
Manage Your Files and Projects Smartly
Use Projects for files you reference repeatedly
If you upload the same document to five different chats, Claude re-reads that document every single time in every session. Use Projects instead. Upload the file once and every new conversation inside that project references it without burning tokens again. Reused project content does not count the same way as fresh uploads, which is confirmed in Anthropic's own documentation.
Keep project instructions concise
Claude reads your project instructions before every single task. If your instructions file runs into thousands of words, that is, tokens burned before any real work starts, every session, every task. Keep instructions under 2,000 words. Include only what Claude genuinely needs for every task. Instructions you only need occasionally belong in a separate file you attach when relevant.
Be specific when using connectors
When connecting Slack, Google Drive, or Notion, vague requests load far more results and burn far more tokens. "Search Slack from the last 7 days for messages about the Q2 launch" is far cheaper than "Search Slack for anything about launches." Filtered retrieval loads only what is relevant and leaves everything else untouched.
Conclusion
Hitting the usage limit is rarely about how much you are asking Claude to do. It is almost always about how conversations are structured. Long sessions, follow-up corrections, full redos, and vague requests quietly multiply your token spend without adding any value to the output.
And if you want practical AI insights like this in your inbox every week, subscribe to the newsletter..
