engineering / skill

Caveman

Skills that ask your coding agent for shorter answers, plus a proxy the README says shrinks what the agent reads, such as logs, test output and web pages.

Repository owner JuliusBrussee

Not rated yet
0 comments

Like 0 Save

Published 11 Oct 2026Listing updated 12 Oct 2026

At a glance

Purpose: Skills that ask your coding agent for shorter answers, plus a proxy the README says shrinks what the agent reads, such as logs, test output and web pages.
Runs in: Chat app with skills, Local agent with shell
Operating systems: Not stated
Needs: Runtime: Node.js 22.13+ for the proxy CLI; Needs internet: Yes; Needs accounts or API keys beyond your agent: Not stated
Files and data: Reads files: Yes; Writes files: Yes; Sends data to other services: Yes
Tool type: Skill
Delivery: From GitHub
Version: Latest on the main branch
Access: Code on GitHub
Compatibility: The skill works in Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline, Copilot and 30 or more others, according to the README. The README shows caveman claude and says the same form takes codex, gemini, aider, kilo, qwen, opencode, hermes, openclaw or pi. It also describes a Claude Code plugin, a Gemini CLI extension, and middleware for the Vercel AI SDK, LangChain, OpenAI and Anthropic calls.
External costs: The repository is Apache-2.0 licensed. The README says the proxy runs with your own keys and your Claude Pro/Max login. It links to Caveman Cloud and states no price. Your coding agent or model provider may charge for usage.
Permissions: The skills act through your coding agent with the access you give it. According to the README, the proxy runs on your machine and keeps the original of each shrunk file on your disk so the agent can pull it back. /caveman-compress rewrites a memory file and backs up the original. caveman learn implement applies fixes one diff at a time, only on your yes. The Claude Code plugin auto-starts every interactive session, subagents too. The README's all-agents installer is a script you download with curl and pipe into bash (on Windows, download and pipe into iex in PowerShell). It covers every agent on your machine and adds a Claude Code statusline badge. The README has you install the command line tool globally with npm, and says the proxy runs with your own keys and your Claude Pro/Max login.
Usage terms: Apache License 2.0, as stated in the repository. See its LICENSE file for the exact terms. The README says older releases and third-party notices are covered in LICENSING.md, and that the Caveman name and rock logo are trademarks of Julius Brussee.

Code on GitHub

JuliusBrussee/caveman

Apache-2.0

Go111,251 stars6,448 forksLast pushed 12 Oct 2026 UTC

🪨 why use many token when few token do trick. Viral skill + proxy for coding agents that cuts 65% of tokens by talking like a caveman.

Topics: ai anthropic caveman claude claude-code llm meme prompt-engineering

Public GitHub data fetched 13 Oct 2026 UTC · 0 days old

View on GitHub ↗

What it does

Caveman is a set of skills and a command line tool for coding agents. According to the README, it has two parts. The skill makes the agent answer in a short, direct voice, and the README says code, commands, paths and error messages stay exact. The proxy shrinks what the agent reads, such as logs, test output, JSON, diffs and web pages.

The README lists the skill modes /caveman, /ultracave and /megacave, plus commands for one-line commit messages, one-line review findings, shrinking memory files such as CLAUDE.md, and showing token use. It also lists work-pattern skills such as investigate-first, lean-build, surgical-patch, safe-refactor, migration and verify-and-stop, and skills that drive the proxy from inside your agent.

The README says the proxy runs on your machine, keeps the original files on your disk, and lets the agent pull them back. It reports benchmark results from its authors and cites outside tests, and it says some cases save nothing, such as HTML pages in its proxy test.

Who it is for

People who use coding agents and want shorter answers and less input to read, and developers who build their own agents.

Example tasks

Switch your agent to caveman mode with /caveman, and say stop caveman to go back.
Write a one-line Conventional Commit with /caveman-commit.
Shrink a memory file such as CLAUDE.md with /caveman-compress.
Run caveman shrink -- pnpm test to compress noisy command output.
Run caveman learn to rank where your tokens go.

Getting started

For the proxy and the skill together, the README gives npm install -g @caveman-ai/cli && caveman setup --install, then caveman claude or the name of another agent. For the skill alone it gives npx skills add JuliusBrussee/caveman -g, and says to type /caveman if it does not start on its own. It also describes a Claude Code plugin route, a Gemini CLI route, an installer script for every agent on your machine, and an uninstall command. The README has the exact commands.

Limitations

The proxy needs Node.js 22.13 or newer, according to the README, and runs on your machine with your keys and your Claude Pro/Max login. The skill alone works in Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline and Copilot. The README says HTML has no compressor yet, so the proxy won nothing back on HTML in its test. It says the skill rules add about 1,160 tokens of text, and that the effect on your bill depends on caching and is not measured yet. It says that if you pay per request instead of per token, for example GitHub Copilot premium requests, a shorter answer costs the same, so you should skip it. It says its authors have not yet tested the Claude app and phone route end to end. The repository can change at any time; check its README before you install.

Data practices

The README says the skill runs on your machine and sends nothing. It says the caveman command line tool sends usage stats by default: commands run, token counts, a random install ID, the OS and the IP, and never your prompts, code or file paths. It says to turn this off with caveman telemetry off or DO_NOT_TRACK=1, and points to SECURITY.md for the full list and for deleting what was sent. The README says the proxy runs on your machine with your keys and keeps every original on your disk. What your coding agent sends to its model provider depends on that agent and provider.