Caveman: Slash AI Coding Agent Output Tokens by ~65% While Keeping Technical Accuracy

Caveman is a plugin (or “skill”) for more than 30 AI coding tools — including Claude Code, Codex, Gemini CLI, Cursor, Windsurf, and Cline — that strips the fluff from agent output while keeping all technical substance. It can shrink model responses by roughly 65% on average, making answers terse, direct, and faster to read without compromising accuracy.

Modern coding agents tend to preface every answer with polite fillers, repetition, and long explanations. This verbosity burns tokens, drains API budgets, slows down iteration, and forces developers to skim through paragraphs of prose to find the fix. Caveman targets the output side of the problem — the “mouth” of the model — and makes it speak in fragments that respect your time and quota.

Caveman offers four modes: lite drops filler but keeps sentence structure; full (default) transforms output into compact fragments; ultra pushes into telegraphic style; and wenyan uses classical Chinese for even shorter rendering. On top of that, sub-commands handle dedicated tasks:

• /caveman-commit writes conventional commit messages ≤50 characters.
• /caveman-review reduces PR comments to one-liners.
• /caveman-stats reports real session token usage and lifetime savings.
• /caveman-compress <file> rewrites memory files (CLAUDE.md, project notes) into caveman-speak, cutting input tokens by ~46% on average.

The installer places a skill file in each agent's known directory. For Claude Code, a hook also writes a small flag file so the agent sees it on startup and starts speaking caveman from the first message. The stats command reads the session log to count saved tokens and feeds that number into a statusline badge. The caveman-compress sub-skill rewrites memory files so every future session starts with a smaller context — savings that compound over time. Importantly, Caveman only changes output tokens; reasoning/thinking tokens are left untouched.

Prerequisite: Node.js ≥18.

macOS / Linux / WSL:

curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash

Windows (PowerShell 5.1+):

irm https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.ps1 | iex

The script detects installed agents automatically; missing ones are skipped. After installation, type /caveman or say “talk like caveman” to activate. Revert to normal with “normal mode”.

Caveman shines during coding sessions, PR reviews, technical Slack threads, and any workflow where output length dominates cost and reading time. It is especially powerful when combined with memory-file compression because both input and output tokens shrink. The plugin does not touch the model's reasoning process, so complex architectural discussions may still require long internal reasoning chains even if the final answer looks short. It also won't help if the bottleneck is input context (unless you use caveman-compress), and the savings vary by prompt type — from 22% on refactoring tasks to 87% on explanatory answers.

Real token counts from the Claude API across 10 representative prompts:

The evaluation harness compares caveman against a baseline of “answer concisely,” not against a verbose default, so the delta is honest.

Caveman is part of a family of tools that apply the same minimal-output philosophy across the development stack:

• caveman-code — a full terminal coding agent built from the ground up for terse interaction.
• cavemem — cross-agent memory that stays compact across sessions.
• cavekit — spec-driven build loops that reduce guesswork.
• cavegemma — a Gemma 4 31B model fine-tuned on caveman-style pairs, baking compression directly into the weights. Together, they form a pipeline where the spec drives the build, the model thinks in compressed form, and every tool speaks only what's necessary.