A tiny Claude Code plugin that strips Claude's filler words — and in one 802-turn session it cut output tokens by 65% and saved ~$118.52. Here's how it works and the real stats.
How a tiny plugin that makes Claude drop its filler words cut my output tokens by 65% — with the real numbers to prove it.
The problem: Claude Code is chatty, and chat costs money
Claude Code is a fantastic pair programmer, but it likes to talk. Every "Great question! Let me walk you through what I'm about to do..." is tokens. Tokens are credits. Credits are money. On a long session — hundreds of turns of building, debugging, and reviewing — all that polite preamble adds up to a real bill.
I wanted the code and the decisions, not the essay around them. That's where Caveman comes in.
What is Caveman?
Caveman is a Claude Code plugin that compresses the model's output into terse, "caveman speak" — stripping out filler words while keeping the actual substance intact. Less "I'll now proceed to refactor this function for you," more "refactor done."
Repo: https://github.com/juliusbrussee/caveman
It runs as a hook on your session, so once it's active, every response comes back compressed automatically. You're not changing what Claude does — only how many words it uses to tell you about it.
Compression levels
Caveman isn't one-size-fits-all. You pick how aggressive the compression is:
| Level | Behaviour |
|---|---|
lite |
Drops filler words only — still readable |
full |
Default caveman mode |
ultra |
Telegraphic, maximum compression |
wenyan |
Classical Chinese style — even shorter |
Switch any time with /caveman lite, /caveman ultra, and so on.
Useful commands
| Command | What it does |
|---|---|
/caveman |
Activate compression for the session |
/caveman-commit |
Write git commit messages ≤ 50 chars |
/caveman-review |
One-line PR review comments |
/caveman-stats |
Show token usage + cost savings |
/caveman-compress <file> |
Compress your CLAUDE.md memory file (~46% smaller) |
That last one is sneaky-good: a leaner CLAUDE.md means less context re-read on every single turn.
How I implemented it
- Add the plugin via the Claude Code marketplace.
- Activate it with
/cavemanat the start of a session. - Pick a level depending on the work —
fullfor everyday building,ultrawhen I just want answers fast. - (Optional) Run
/caveman-compress CLAUDE.mdto shrink the memory file.
If the install ever breaks, the trick is to just tell Claude Code: "Read CLAUDE.md and INSTALL.md, install caveman for me." The agent fixes its own brain.
The real results
Here's the part that actually matters — numbers from one of my own sessions, pulled straight from /caveman-stats (priced on claude-opus-4-8):
| Metric | Value |
|---|---|
| Turns in session | 802 |
| Actual output tokens (with Caveman) | 850,895 |
| Estimated output tokens without Caveman | 2,431,129 |
| Tokens saved | 1,580,234 (~65%) |
| Estimated cost saved | ~$118.52 |
That's an 802-turn session — a serious chunk of real work — where Caveman cut output token usage roughly in two-thirds and saved over a hundred dollars. On a single session.
The honest take: what Caveman actually saves
A fair caveat so nobody gets the wrong idea: Caveman compresses output tokens — the words Claude generates back to you. It doesn't shrink your input context or the cache-read tokens that get re-read each turn (in this session, those cache reads were 258M+ tokens on their own). So the savings land on the generation side, not the whole bill.
It's also worth noting Caveman's own published benchmarks are more modest than my session suggested — savings vary a lot by task. My 65% was a long, verbose, explanation-heavy session, which is exactly where compression pays off most. Tighter, code-only tasks will save less.
Trade-offs
- Readability.
ultraandwenyanare terse. Great for speed, less great when you actually want Claude to explain its reasoning. I keep a level that matches the task. - Onboarding others. If a teammate reads your session logs, caveman speak takes a second to parse.
Verdict
For long, iterative building sessions where I already know what I'm doing and just want the work — Caveman is an easy win. 65% fewer output tokens and ~$118 saved in one session is not a rounding error.
My rule of thumb: full for daily driving, ultra when I'm moving fast and confident, drop back to normal mode the moment I want Claude to actually teach me something.
Less words. Same code. Lower bill.