Caveman: make Claude talk like a caveman to cut output tokens
TL;DR
Caveman is a community skill/plugin for Claude Code that makes Claude reply in terse, telegraphic style. Generated code is unchanged; only the prose is cut. The project cites about a 65% cut in output tokens, but total session cost drops much less because output is only part of it.
Claude usually answers politely and thoroughly: a greeting, a lead-in, step-by-step explanation, a final summary. When you already understand the work, most of those words just cost output tokens and reading time. Caveman is a community skill/plugin for Claude Code that forces Claude to answer extremely tersely, telegram-style (no articles, no pleasantries), while the generated code stays intact.
What it does, and doesn't do
- It only shortens the explanations; code, commands and file paths are untouched.
- The project reports about 65% fewer output tokens.
- Output is only part of a session's cost (input and context are usually larger), so the overall saving is much smaller than that figure.
Modes
| Mode | Description |
|---|---|
| lite | Drops filler, keeps full grammar |
| full (default) | Clipped sentences, no articles |
| ultra | Maximum compression, note-style |
There are also variants for short commit messages, one-line PR reviews and compressing CLAUDE.md files to cut input tokens.
Installation
claude plugin marketplace add JuliusBrussee/caveman
claude plugin install caveman@caveman
# then type /caveman in the sessionWhen to use it, when not to
| Use | Don't use |
|---|---|
| Repetitive work you already understand (small fixes, familiar refactors) | Learning a new framework, when you need explanations |
| Long sessions with many exchanges | Architecture decisions, debugging strange bugs |
| Commit messages, internal notes | Content sent to colleagues or customers |
Risks
- Losing the "why": overly clipped answers can drop the reasoning, making later review harder.
- Misunderstanding: with vague requests, terse answers can hide the fact that the agent guessed.
- Community dependency: this isn't an official Anthropic product; check for updates and read the skill before installing.
The no-install way
Add a short section to CLAUDE.md:
## Communication style
- Answer concisely, no greetings, no recap of what you did.
- Only explain when I ask "why" or when there's a risk.This costs a few dozen input tokens per session and gets most of the benefit for ordinary users.
How to apply it
- Dev: turn it on for long refactoring sessions; turn it off when you need explanations or design.
- Tech Lead: agree a response style in the repo's CLAUDE.md instead of letting everyone install their own.
- PM: saving output is a small optimisation; don't prioritise it over bigger levers like splitting sessions and keeping context lean.