caveman skill · Short Agent Replies (JuliusBrussee)

What it is, and what it is not

caveman is a voice, not broken grammar. skills/caveman/SKILL.md tells the agent to answer first, cut ceremony (“Sure!”, “Let me”, “Hope this helps”), keep one idea per sentence and drop articles only when the sentence still reads in one pass. not, never, no and only are never dropped. Code blocks, commands, paths and error strings are copied character for character. Security warnings, irreversible actions and a confused user get full sentences, and so does anything written outside the chat: code comments, commits, docs and PRs.

The same repo also ships a much larger caveman CLI. The skill is one part of it; the README calls the parts “rocks”. Pick the one you mean:

The full install is npx @caveman-ai/cli@latest. It finds your agents and logins and shows what it will change before you press Enter. This page covers the skill; read the README before you run the CLI.

Same name, different package: skill collections such as alirezarezvani/claude-skills ship their own “caveman” skill. This entry covers JuliusBrussee/caveman only.

Install for your assistant

Commands follow the JuliusBrussee/caveman README and INSTALL.md, fetched 2 Nov 2026. AgentSkillsDir has not run them yet.

• skills CLI (skill only)

  then type /caveman if it does not start
  npx skills add JuliusBrussee/caveman -gCopy

  The skill alone, no proxy. -g installs into your user skill folder; drop it to install for this project only. Pick an agent with -a (codex, windsurf, cline, kilo …). Remove with npx skills remove JuliusBrussee/caveman -g.

• Claude Code (plugin)

  then restart; it auto-starts every session
  claude plugin marketplace add JuliusBrussee/caveman claude plugin install caveman@cavemanCopy

  Auto-starts in every interactive session, subagents too. Already inside Claude Code? Type /plugin marketplace add JuliusBrussee/caveman, then /plugin install caveman@caveman. Update with the full name: claude plugin update caveman@caveman (plain “caveman” fails). Auto-start needs Node.js 18+ on PATH; without it, type /caveman.

• Codex (skills + start hook)

  then run /hooks once, trust the caveman hook
  npx -y github:JuliusBrussee/caveman -- --only codexCopy

  Installs the skills and a SessionStart hook, so every new Codex session, /clear or compaction starts in caveman. Skills only, no hook: npx skills add JuliusBrussee/caveman -a codex -g, then type $caveman (Codex calls skills with $, not /). On npm 12+, add --allow-git=root after npx.

• OpenCode

  then new session; it is on by default
  npx -y github:JuliusBrussee/caveman -- --only opencodeCopy

  Adds a native plugin plus the always-on rules in AGENTS.md. Subagents read those fixed rules, so “stop caveman” does not switch them off for subagents. On npm 12+, add --allow-git=root after npx.

• Cursor (session hook)

  then new chat starts in your default mode
  npx -y github:JuliusBrussee/caveman -- --only cursorCopy

  Adds the skills, the Cavecrew agents and a session hook (~/.cursor/hooks.json; your other hooks stay). The alternative is the Cursor plugin: clone the repo into ~/.cursor/plugins/local/caveman and reload the window. Pick one; with both, every chat gets the rules twice.

Other agents

Gemini CLI

gemini extensions install https://github.com/JuliusBrussee/caveman
Auto-activates.

Windows (PowerShell 5.1+)

irm https://raw.githubusercontent.com/JuliusBrussee/caveman/v4.0.1/install.ps1 | iex
Every detected agent at once. If execution policy blocks it: Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass, then re-run.

Windsurf, Cline, Kilo, Roo, Amp …

npx skills add JuliusBrussee/caveman -a windsurf -g
Swap windsurf for your agent’s profile. Per-session: type /caveman or ask for caveman mode.

Claude apps (web, desktop, mobile, Cowork)

Customize > Plugins > Add marketplace: JuliusBrussee/caveman
No startup hook in chat: start with /caveman. Upstream has not tested this end to end.

Full matrix (30+ agents), dev containers, update and uninstall steps: INSTALL.md.

Turn it on, switch modes, turn it off

1. Type /caveman ($caveman in Codex) if it did not start on its own. Ask a real question: the reply should open with the answer, no greeting or recap.
2. Want it shorter? /ultracave keeps only fragments. /megacave answers in Classical Chinese (文言); upstream says its token savings are small and noisy.
3. Say stop caveman or normal mode to go back. /caveman status shows the current mode and changes nothing.

Mode is per window. To pick the mode new sessions start in, put {"defaultMode": "ultracave"} (or caveman, megacave, off) in ~/.config/caveman/config.json, or a .caveman.json at a project root. Scripted Claude Code runs (claude -p, the Agent SDK) start with caveman off. Not yet verified by AgentSkillsDir.

About the savings numbers

Upstream’s own baseline is “Answer concisely”, which new models already follow. On ten dev questions, /caveman wrote 3% fewer output tokens than that at the median and /ultracave 35% fewer. These were measured on the older 3.1.0 rules and not yet re-run. The rules cost about 1,160 input tokens, and if you pay per request rather than per token (GitHub Copilot), shorter answers save nothing. JetBrains’ paired test on 86 tasks found 8.5% fewer output tokens and no measurable quality loss.

The bigger percentages on the repo page belong to the full CLI or to one example, not to the skill on your work. They are upstream claims, not reproduced by AgentSkillsDir. Read every case where caveman loses and measure on your own sessions.

caveman and ponytail

Both are behaviour skills, and they stack. ponytail decides how much code the agent writes; caveman decides how the reply reads. Upstream INSTALL.md says both apply when installed together. Want only one in a window? Say stop caveman there. graphify is a different job again: it maps a codebase into a graph the agent can query.

Before you install

• SKILL.md is plain written instructions. Read it on GitHub first.
• The plugin and installer add session hooks that run Node, and edit agent config files (settings.json, hooks.json, AGENTS.md). Run the installer with --dry-run to see every command first.
• The full CLI sends usage stats by default and Auto sends your asks to Caveman Cloud once you pick it. See SECURITY.md. The skill alone sends nothing.
• Listing is not a security review. AgentSkillsDir has not reviewed this code.

FAQ

Related

• Browse all Claude skills in the directory
• ponytail skill: the smallest complete code change
• Read the official INSTALL.md on GitHub

Corrections and suggestions

Spotted something out of date, or want to suggest a skill? Email a correction. To suggest a resource, submit it for manual review.