Stop Burning Claude Tokens: Tips From Someone Who Learned the Hard Way
Here are some tips to reduce Claude token usage and keep things efficient, based on my experience and some things I've been experimenting with. Some of these sound obvious until you realize you've been doing them wrong for weeks.
Long chats: Avoid dragging out conversations when the previous context isn't needed. Use /clear and start fresh. Claude processes the entire conversation from scratch on every response, so long chats = more tokens every time. Yes, every single time.
Don't let the agent read entire files: This is probably the biggest waste. Share only the relevant section, or tell it exactly which function to look at, or specify a line range (e.g., "read from line 40 to 80"). Letting Claude read a 2000-line file to fix one function is like calling a plumber to fix a leaky tap and making them inspect the entire building first.
Be specific with what you want: Instead of "Emoji not showing on all phones, fix that", say "Emoji screen showing black lines instead of emojis when the user selects a custom font." The more specific, the better. Vague prompts = Claude guessing = you correcting = more tokens wasted.
Start a new chat for a new task if it doesn't need context from the previous one.
Don't always default to the highest model: Save the powerful models for tasks that actually need deeper thinking. Not every bug fix needs the smartest brain in the room.
Edit your last message instead of sending a follow-up correction when Claude misunderstands. Sending "no not like that, I mean..." just burns extra tokens. Just edit the message, it's right there.
Use /context to see what's actually eating up your context window. Eye-opening honestly.
Put persistent rules and instructions in CLAUDE.md so you're not repeating yourself every session. If you're typing the same instruction more than twice, it belongs in CLAUDE.md.
Start in plan mode to scope out features or implementation details before diving in. It saves you from Claude going full speed in the wrong direction.
Avoid thinking/ultrathink mode unless the task actually requires it. Great feature, expensive feature.
Only enable the MCP servers you actually need. Disable the ones you're not using. No reason to bring the whole toolbox when you just need a screwdriver.
Use images instead of describing UI issues: Claude Code accepts drag and drop images in the CLI and it really helps for UI bugs. Just don't re-upload the same image multiple times, each upload is treated as a separate file.
Prevent Claude Code from generating unnecessary docs and summaries. If that's not part of your workflow, add it to CLAUDE.md.
Set Claude Code to respond concisely and directly. Shorter responses = fewer tokens. Nobody needs a five-paragraph essay about why it changed one line.
For bigger tasks, check what Claude is actually thinking during its thinking state. If it seems like it's overcomplicating things or missing the point, edit your previous message to steer it before it goes too far down the wrong path.
When you manually edit the code, let Claude know. Otherwise it won't be aware of your changes and may revert them or make mistakes based on outdated context. This leads to a painful back-and-forth just to get it back on track.
Provide the library version you're using: Claude will look up the exact docs for that version instead of guessing or defaulting to whatever version it feels like. Small thing, big difference.
Share the relevant library docs or function reference directly: If it's easy to grab, paste it in. Saves Claude from hunting it down itself and makes sure it's working with the right information.
Currently Experimenting
- When a task requires full context and understanding of something, ask Claude to first read and understand the full flow, then write a summary of it to an .md or .txt file. For future tasks that need that same context, just tell Claude to refer to that file instead of re-reading everything from scratch. Still testing this but it's looking promising.
