Cut Claude Code Token Usage by 70 Percent with 11 Fixes

The screen that opens my latest video is a Claude Code session at 100 percent used, with three and a half hours still to go before the reset. The screen it closes on is a session at 20 percent used with 30 minutes left. Eleven changes sit between those two screens, and together they cut my usage by about 70 percent.

I assumed the model was the expensive part. Mostly it was not. What was eating the window was noise: the output of commands nobody needed to read, a CLAUDE.md file read again on every turn, and tool definitions from servers I never used. This guide walks through all eleven, in the order I would do them.

Run /context before you change anything. It shows exactly what is filling your window, with an estimate for each category: system prompt, system tools, custom agents, skills and so on. The usual culprits are a fat MCP chunk and a bloated brief.

The point of doing this first is that the expensive thing is rarely the one you would guess. Two minutes with /context tells you which of the changes below is worth your time on your setup, which beats working down somebody else's list in order.

Two of the eleven are tools you add. They attack the problem from opposite ends, one on what goes into the context window and one on what comes back out of it.

rtk, for what goes in

rtk sits between Claude Code and your terminal. Every command's output passes through it, and it hands Claude back a clean version with the unnecessary data stripped out. In the example in the video, a command that came back as 2000 tokens comes back as 400, which is about 80 percent.

On its own, rtk can save 60 to 90 percent of your token usage. How much depends on your use case, because noisy commands save the most and not every command is noisy. It removes noise rather than signal, so there is no loss in the quality of what Claude does with the result.

caveman, for what comes back

caveman does the same job in the other direction. It makes Claude Code explain only the important parts instead of writing a wall of text, and it does that without changing the quality of the code or the agentic work.

This one deserves an honest number rather than the best one. On one task the output went from 1180 tokens without caveman to 159 with it, around 87 percent. Other tasks only saved 41 or 30 percent. It helps, and how much it helps depends on what you ask for.

The rest are habits, and every one of them is already built into Claude Code. /context above is the first of the nine. Here are the other eight.

• Run /compact yourself at 60 percent. /compact summarises the session and keeps only what matters. Left alone, Claude Code compacts automatically at 100 percent, which breaks the cache in the middle of a session, after quality has already started to drop. Do it yourself at a clean stop. In the video one compact took the window from 60 percent to 13.
• Hand a long session over to a fresh one. Every message you send makes Claude Code read the whole chat again. Ask it to write a handoff prompt covering what has been done and what still needs doing, then start a new session with that prompt. You keep the context and drop the weight.
• Pick one model and stay on it. Reused context is cached at about a tenth of the price. Changing models in the middle of a session makes it consume more tokens, at full price. Choose your model with /model at the start.
• Keep CLAUDE.md under 200 lines. It is read on every message, so every line in it is a cost you pay again and again.
• Feed it markdown, not PDF. Large noisy files like PDFs cost far more than the same content as markdown.
• Disconnect the MCP servers you never use. Every connected server injects its tool definitions into every message, even the ones you never touch. Run /mcp to see what is connected.
• Plan before it writes a line. Start sessions with /plan. You get a full plan to accept or edit before any code is written, which stops the wrong attempts Claude Code makes in auto mode.
• Set /model opusplan. Opus writes the plan, and once you approve it Sonnet writes the code. With the plan already mapped, the gap between the two models on the writing itself is small, and you stop paying Opus prices for every line.

If you only do one thing today, run /context. If you do two, install rtk. Everything else can wait for your next session, because none of it needs anything new.

The CLAUDE.md file with all of this already set up is free in my Skool community: get the file. The full walkthrough, with the before and after screens, is in the video above.