01How to Reduce Claude Tokens: Why Your Usage Disappears
Learning how to reduce claude tokens starts with one fact: Claude resends your whole conversation with every message, so the words you type are only a tiny slice of the total. Stacking free settings and simple habits can cut your usage by up to 80%, depending on how you work today.
Most people write shorter messages and ask fewer questions. Julian Goldie explains in this episode why that barely moves the needle. Claude has no memory between messages, so each new one carries your first message, every reply, every file it opened and every long command result.
Each time Claude uses a tool, it sends another request carrying all of that again. The pile only ever grows. A one line question in a session that has been open all day still draws on your usage for the whole day's conversation.
There is also a part you never see. More than 60,000 tokens can load before you type a word.
02Fix 1: Upload Your Files Instead of Pasting Them
The fix that surprised Julian most has nothing to do with coding. It is simply how you give Claude a file, and it works in the Claude app too. You do not need to be technical to use it.
In a test using a table of data, the two versions got the same data and the same questions:
- The pasted version scored 6 out of 25.
- The uploaded version scored 25 out of 25.
- The uploaded version used about a 12th of the tokens.
So this fix gave better answers and much lower usage at the same time. It also answers a common doubt, which is that Claude might get worse if it sees less. Less junk often means better answers.
The habit is easy to build. When you are about to paste a spreadsheet straight into the chat, attach the file instead. This one change needs no setup and no special tools, which makes it a good place to start.
03Fix 2: Stop Giant Command Results Flooding Your Chat
This is the biggest hidden drain in Claude Code. When Claude runs a command like a test run, it reads every line that comes back. A test run can print thousands of lines when Claude only needed the one that failed, and those lines get reread on every message after that.
Anthropic's own docs give a fix. A hook is a small rule Claude Code runs automatically at a certain moment. In their example, a hook pulls only the error lines from a 10,000 line log file, taking it from tens of thousands of tokens to hundreds.
Julian built this into a mod called Token Saver, which is a small add-on for Claude Code. Here is what it does:
- Keeps the top, the bottom and every line that mentions a problem, plus the few lines after each one.
- Saves the full result to a file and tells Claude where it is.
- Stops Claude the first time it tries to read a huge file and tells it to search instead.
- Shows a live meter and warns you when your context passes 60%.
One noisy command went from about 71,000 tokens to about 1,400, and the failure and its reason were still there. Your whole session will not drop that much. The mod comes with install steps inside the AI Profit Boardroom.
04Fixes 3 and 4: Turn On Tool Search and Trim Your Rules File
Every tool you connect has a description that takes up room. The fix is called tool search. Claude keeps only the names and looks up the details when it needs one, and it also got better at picking the right tool.
On Julian's own machine, the same short message went from about 89,000 tokens to about 61,000 with one setting changed. Claude now holds these details back by default, so most people already have this. Your job is to clean up what is left:
- Type /context to see what is taking up the room.
- Type /mcp and turn off anything you are not using today.
- If you run Claude through a gateway or a proxy, check tool search is on, because it can switch itself off without telling you.
Next, look at your rules file, and keep it under 200 lines. If it describes your business and your project folders, Claude gets that straight away instead of opening 10 files to work it out.
05Fix 5: How to Reduce Claude Tokens With /clear and /compact
This is the easiest habit on the list. When you finish one job and start something unrelated, type /clear. All that old context would otherwise be reread on every message, and /clear itself uses nothing.
If you are halfway through something, use a handoff note:
- Every couple of hours, ask Claude for a three line note on what you are doing, what you tried and what is next.
- Copy the note.
- Type /clear.
- Paste the note back in.
Token Saver has a /handoff command that writes this prompt for you.
To keep the conversation going, use /compact, which summarizes it. Do this yourself at around 60%, at a natural break such as when research ends and building starts. Do not wait until you hit the wall.
Compacting has to read everything it summarizes, so doing it on a nearly full session is a big request. You can also tell it what to keep, such as the code changes and the test results.
Breaks matter too. Claude keeps recent conversation in a cache that is much lighter to reread. Walk away long enough and it goes cold, so your next message reprocesses everything.
06Fixes 6 and 7: Choose the Right Model and Effort Level
You do not need your biggest model on the highest effort for everything. Sonnet handles most jobs, so set it as your default. You can switch with /model right in the middle of a session.
The strongest version of this is to plan with the big model and then build with the smaller one. Plan mode lets Claude look around and write the blueprint before it touches anything.
In Julian's earlier token lesson, routing work like this cut usage by around half. Planning first and building on a lighter model came out around 59% lower than using the big model for both.
Fix 7 is the effort setting, which is how hard Claude thinks before it answers. All that thinking counts as output, and output counts heavily. Julian gave Claude the same prompt at three levels, and the same three sentences came back each time:
- Low wrote 108 output tokens.
- High wrote 358 output tokens.
- Extra high wrote 963 output tokens.
Type /effort and match it to the job. Use low or medium for a blog post or tidying a list, high for a real problem, and the very top setting only when a job truly needs it.
07Fixes 8 and 9: Move Noisy Work Away and Give a Clear Target
Running tests, reading docs and digging through logs all fill your conversation up fast. Hand those jobs to a sub agent, a helper that works in its own separate conversation.
Be careful with agent teams, though. Anthropic says they use about seven times more tokens than a normal session when teammates run in plan mode, because each one has its own full conversation. Keep teams small, give each one a tight job and shut them down when they finish.
Fix 9 is to give Claude a clear target. A vague request like improve this project makes Claude go and read everything. A specific one, like adding a check to the login form in one named file, lets it go straight there.
- Tell it what done looks like, such as a test that should pass or a screenshot of the result.
- If it heads the wrong way, press escape straight away.
- Then type /rewind to go back to a good point.
When Claude can check its own work, it catches problems before you need another round.
08Fix 10: Free Token Tools Compared With Real Numbers
Several free tools promise big savings, but the numbers on the box do not always hold up. Here is how they looked in this episode:
- Headroom shrinks what Claude reads. It reports 15 to 20% on coding and 60 to 95% on structured data, so it is worth using if you work with a lot of data files.
- Ponytail makes Claude act like a lazy senior developer, so it builds less. It averages about 54% less code, and one set of 80 paired runs came out about 10% lower per task.
- Caveman makes Claude talk in short sentences. It advertises 65%, but its own table shows about 8.5% on real coding jobs and about 50% on short questions.
- RTK filters command results. Its own benchmark cut them by 77%.
RTK is the surprise. When JetBrains tested it on real agent work at low effort, total usage went up by about 7.5%. The filtering worked, but with less to read, the agent sometimes took a longer route and needed more steps.
The lesson applies to Token Saver too. Less text going in does not always mean less used overall. So install one thing, do a normal job and compare.
Keep it if your number drops, and take it off if it does not.
09How to Measure Your Savings and What to Do This Week
Type /usage to check your numbers. On a paid plan, it breaks usage down by skills, helpers, plugins and connections and gives suggestions. On newer versions, you can press F to have Claude Code apply the ones it can. /insights goes further and writes a report on how you work.
Also stop any scheduled loops you do not need, because a loop sends your full context every time it fires, even while you are away. Keep web search off until you need it.
This is for anyone who keeps getting stopped halfway through their best work, and most of it only takes typing a short command. If you run one giant chat on the top model, you have a lot to gain. If you are already careful, you will save less, because 80% is the top of the range.
- Type /usage and /context, and write down what takes the most room.
- Switch off connections you do not need, and cut your rules file under 200 lines.
- Use /clear between jobs and upload files instead of pasting them.
- Set your default to Sonnet, lower your effort for simple work and install Token Saver.
Then do a normal job and check your number again.