This article is one chapter of my book Delphi in all its glory – AI-assisted development, the fifth book of the series.
Claude is verbose. Ask it to fix a one-line bug and you get three polite paragraphs before the fix arrives. Every one of those words is an output token, and output tokens cost five times what input tokens cost.
CaveMan is a plugin that fixes this by making Claude talk like a caveman. It drops articles, filler and pleasantries, keeps every technical term exact, and leaves code blocks untouched. “I would be happy to help. The issue is likely caused by your token expiry check” becomes “Token expiry check use < not <=. Fix:” followed by the same code. It installs in one line:
claude plugin marketplace add JuliusBrussee/caveman claude plugin install caveman@caveman
Julius Brussee wrote the first version as a joke, in ten minutes. It now has over 100,000 GitHub stars, which tells you how much AI verbosity annoys people. Watch the license if that matters to you: the skill is MIT, but the engine is under a Business Source License that only turns into Apache-2.0 in 2030.
Now the honest part, because the headline number is oversold. “65% fewer tokens” is real, but it is output tokens only. In a real session, output is the small half. Most of your tokens are input: your CLAUDE.md, the units Claude reads, the conversation so far, every tool result. Measure the whole session and the saving lands closer to a quarter.
The quality argument is better than the cost one, and it is not just marketing. A March 2026 paper, “Brevity Constraints Reverse Performance Hierarchies in Language Models” (arXiv:2604.00025), tested 31 models across five datasets and found that forcing a large model to answer briefly improved its accuracy by 26 percentage points. Big models talk themselves into mistakes. Making one shut up and get to the point can make it more correct, not less.
So why am I not using it? Because I do not only write code with Claude. I write documentation, emails, and this book. CaveMan has to be turned off for every one of those and turned back on afterwards. You can automate the switch with a rule in CLAUDE.md, but that is one more rule to maintain, and a rule Claude sometimes ignores.
The verdict: if your sessions are pure coding, install it and enjoy the reduced scrolling. If you also write prose with Claude, you will spend more attention on the toggle than you save in tokens.
This was one chapter of a whole book
You just read one chapter of Delphi in all its glory – AI-assisted development: more than 400 pages about Claude Code and AI for Delphi programmers. Written by a Delphi programmer, with the failures documented next to the wins.