Caveman
JuliusBrussee/caveman
Caveman is a free, Apache-2.0 skill pack by Julius Brussee that makes a coding agent drop filler words and answer in short fragments, while code, commands, file paths and error messages stay as they are. The repo also ships an optional local proxy and a middleware library that shrink the tool output an agent reads. The README says the skill works in more than 30 agents.
Facts checked on against the sources listed at the end of this page.
- GitHub stars
- 114,429
- Skills in the pack
- 22
- Agents supported
- 9plus any agent the skills CLI supports
- License
- Apache-2.0
- Last commit
- 5 days ago17 October 2026
- Instruction size
- 11.2k wordsacross all SKILL.md files
How to install Caveman
| Agent | Command | Copy |
|---|---|---|
| Any agent the skills CLI supports | Type /caveman if the agent does not switch on by itself. | |
| Claude Code | | |
| Codex | | |
| Cursor | Run the Codex command and replace codex with the Cursor agent profile, as the README says for Codex, Cursor, Windsurf, Cline and other skills-compatible agents. | |
| Gemini CLI | |
Commands for 1 more agent
| Qwen Code | |
|---|
Commands as given in the README on 2 October 2026.
What Caveman runs on your machine
- Hooks
- Yes, with the full installer. It wires Claude Code hooks and a statusline badge. The repo's .codex/hooks.json adds a session-start hook that prints the caveman rules. The skill alone, installed with npx skills, adds no hooks.
- Bundled scripts
- Yes. The repo ships a shell installer (install.sh) and a PowerShell installer, plus a command-line tool and a local proxy written for Node.js 22.13 or later.
- Network calls
- The skill alone: none, per the README. The caveman CLI sends usage stats by default, including a random install ID, OS, CLI version, timezone and IP address, but never prompts, code or file paths. caveman telemetry off or DO_NOT_TRACK=1 turns it off.
- API keys needed
- None for the skill. The README says no account or API key is needed. The proxy passes your own provider login or key through to your provider.
These are facts read from the repo, not a security rating. Read the files before you install any skill.
What is inside Caveman
| Skill | What it makes the agent do |
|---|---|
| caveman | Sets the talking style at lite, full, ultra or wenyan intensity. Code, commands and error messages are not shortened, and security warnings use full sentences. |
| caveman-commit | Writes a terse Conventional Commit message. |
| caveman-review | Gives one-line, actionable review findings. |
| caveman-compress | Shrinks a Markdown memory file such as CLAUDE.md, keeps headings, paths and commands, and backs up the original. |
| caveman-stats | Shows recorded Claude Code token usage for the session. |
| cavecrew | Compressed subagent presets: investigator, builder and reviewer. |
| caveman-learn | Reads your local agent history and ranks where your tokens go. |
| lean-build and surgical-patch | Work patterns that make the agent write less code for a task. |
| investigate-first and verify-and-stop | Work patterns that read before editing and stop once a change is verified. |
| safe-refactor and migration | Work patterns for refactors and migrations. |
The remaining skills drive the optional proxy: caveman-setup, caveman-discover, caveman-manage, caveman-optimize, caveman-explore, caveman-evidence-review and caveman-help.
A good fit for
- Developers who read agent answers more than they paste them, and want shorter replies
- Long sessions full of logs and test output, where the optional proxy trims what the agent reads
- Teams that pay per token and want to measure savings with the repo's own trial command
- People who want to shrink a large CLAUDE.md or other memory file with /caveman-compress
Skip it if
- Plans billed per request, such as GitHub Copilot premium requests, where a shorter answer is still one request
- Work that is almost all code generation, because the rules add about 1,000 estimated input tokens to every call and the README says that can cost more than it saves on terse answers
- Anyone who wants no telemetry by default, since the CLI sends usage stats until you turn it off
How Caveman works
The core skill is one rule file. It tells the agent to drop articles, pleasantries and hedging and to answer in fragments, at an intensity you set with /caveman lite, full, ultra or wenyan. It will not shorten code, rewrite an error message or grunt through a security warning, and it returns to full sentences for anything irreversible.
The same repo adds two larger parts that you can skip. A local proxy sits between your agent and the provider and compresses logs, JSON, diffs and test output, keeping byte-exact originals in a local SQLite file. A middleware package does the same inside an app you build. The README reports a 33.2 percent drop in input tokens on a pinned 54-run benchmark, and cites a JetBrains test of the skill alone that found 8.5 percent fewer output tokens with no detectable quality change.
- Made by
- Julius Brussee
- Topic
- Coding workflow
- Works with
- Claude Code, Codex, Cursor, Gemini CLI, Windsurf, Cline, GitHub Copilot, Qwen Code, Oh My Pi
How much context Caveman uses
| All 22 skills | 11,166 words across every SKILL.md file |
|---|---|
| Largest skill | caveman-learn, 1,891 words |
| Loaded at start | Only each skill's name and description. The full text loads when a task needs that skill. |
Word counts are measured from the repo every day. Reference files a skill loads on demand are not counted.
Is Caveman still maintained?
| Last commit | , 5 days ago |
|---|---|
| Latest release | v4.0.1, 13 October 2026 |
| Stars gained in 7 days | +1,800 |
| Stars gained in 30 days | Shown after 30 days of tracking |
| Open issues and pull requests | 87 |
| Forks | 6,595 |
Tools Caveman works with
Similar skill packs
Questions about Caveman
Does Caveman hurt answer quality?
The README cites a JetBrains test of 86 coding tasks that found 8.5 percent fewer output tokens and no detectable quality change for the skill alone. It also says its own 10-question eval measured length, not correctness.
Do I need the proxy?
No. The skill works on its own with npx skills add JuliusBrussee/caveman -g. The proxy and middleware are separate, optional parts that shrink what the agent reads.
Does it change my code or commands?
No. The README says code, commands, file paths and exact error messages are never cavemanned, only the prose around them.
Is it free?
Yes. The whole repo is Apache-2.0 from version 3.0.0, and releases before that keep their original license. The README mentions a separate hosted product, Caveman Cloud, with a waitlist.
How do I turn it off?
Say stop caveman or normal mode, or use /caveman off. To remove the install, run npx -y github:JuliusBrussee/caveman -- --uninstall.
Sources
Something wrong or out of date? Email [email protected] with a link to the source. Stats come from the public GitHub API and update every day.
For the maker
Add the badge to your README or site. The number updates every day.
Sharing on X? Open the share card with the pack's live numbers.
