Decent article on how to query claude without burning too many tokens.
Summary:
1. Edit instead of following up
Go to your prompt → click the pencil icon → rewrite it → hit send. Claude regenerates as if the bad response never happened, so you avoid piling up conversation history.
2. Start fresh every 15–20 messages
When a chat gets long, ask Claude "Summarize everything we've discussed so far," copy that summary, open a new chat, paste it as your opening message, and continue there. Avoids the exponentially rising token cost of long threads.
3. Batch questions into one message
Instead of three separate asks, combine them — e.g. "Summarize this article, list the main points as bullets, then suggest a headline" — so context is only loaded once instead of three times.
4. Upload recurring files to Projects
Sidebar → Projects → create a new Project → upload the files you use repeatedly. They get cached, so you're not paying token costs to re-upload/re-read them every conversation.
5. Set up Memory
Settings → Memory → add personal details (job, writing style, preferences). Or alternatively, keep a .md file with your context to paste in, so you're not re-explaining yourself each time.
6. Turn off unused features
Click + → turn off Research, web search, or connectors you're not using. Click the model selector → turn off extended thinking if you don't need it. Each active feature adds token overhead even when idle.
7. Match model to task
Click the model selector at the top of the chat. Use Haiku for quick/simple tasks, Sonnet for everyday work, Opus only for genuinely complex analysis — cheaper models stretch your limit further.
8. Spread work across the day
Split usage into 2–3 sessions per day instead of one long session. Since limits run on a rolling 5-hour window (not a daily reset), earlier usage ages out and frees up capacity between sessions.
Core takeaway: "Claude counts tokens, not messages" — usage patterns matter more than message count.
