Caveman skill: terse, answer-first replies from Claude Code and other agents

Source: JuliusBrussee/caveman · Maintainer: community: Julius Brussee · License: Apache-2.0 · Checked: 24 October 2026

What is the caveman skill?

The caveman skill makes your coding agent reply in a terse, answer-first style: no greeting, hedging, recap or sign-off, short words and fewer articles, while code, commands, file paths, numbers and error messages stay character for character. It is for developers who read agent output in a terminal and want less to scroll through.

It comes from JuliusBrussee/caveman, whose README tagline is "why many token when few do trick." The skill's description reads: "Terse caveman voice: answer first, fluff gone, every technical fact kept." The SKILL.md opens with "Caveman is a voice, not broken grammar."

Its skills.sh page listed about 590,000 installs (snapshot, 24 October 2026), with Pass results from Gen Agent Trust Hub, Socket and Snyk. The repo had about 115,000 GitHub stars on the same day.

What does it do?

The SKILL.md sets nine rules and a list of times to break them. It persists: "Every response, whole session, until user says "stop caveman" or "normal mode"."

The rules, in short:

1. Answer first. Pattern: [thing] [action] [reason]. [next step].
2. Kill ceremony. No "Sure!", "Let me", "Hope this helps", and no just, really, basically, actually or simply.
3. Short words. "fix" not "implement a solution for." Standard acronyms are fine; invented abbreviations like cfg or impl are not.
4. Articles optional, meaning never. a, an and the can go; not, never, no, only and except never do.
5. One idea per sentence. It uses ASD-STE100, the controlled English of aircraft maintenance manuals, as the floor: 20 words max, active voice.
6. Payload verbatim. Code blocks, commands, paths, API names and errors stay exact, and existing comments in files it edits are never shortened.
7. Bounded status during tool runs. No text between routine calls; one line per phase and one with the result.
8. Your language. It compresses style, not the language you write in.
9. Never perform caveman. No "me think," no "Caveman:" prefix, no emoji, and "Caveman phrasing not shorter than plain? Use plain."

When it switches back to plain prose: security warnings, irreversible actions, step order a fragment could scramble, a confused user, questions it asks you, and "Anything persisted outside chat: code, comments, commits, docs, issues, PRs, tickets, memory files, third-party messages."

The repo's own example, for "Why does my React component re-render?":

New object ref each render, so React re-renders. Wrap the prop in useMemo.

Siblings. caveman is the default of three voice skills. ultracave strips grammar further, and megacave answers in classical Chinese with technical terms kept verbatim. /caveman ultra and /caveman wenyan are aliases that hand off to those skills, and /caveman status reports the current mode.

The caveman folder holds only SKILL.md and a README.md, with no scripts. The plugin route adds hooks that switch the mode on at session start.

How do you install it?

The repo ships far more than this skill: a CLI, a proxy that compresses what the agent reads, model routing and more. The README offers a skill-only path: "Only want shorter answers? Get just the skill."

Skill only, all agents the skills CLI finds:

npx skills add JuliusBrussee/caveman -g

That command points at the whole repo, which holds many skills. To take only this one, skills.sh shows:

npx skills add https://github.com/juliusbrussee/caveman --skill caveman

If you want the /caveman ultra alias to work, install ultracave too, since the alias follows that skill. The skills CLI accepts --skill more than once.

Claude Code plugin. The README says this one "auto-starts every interactive session, subagents too":

claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman

See Claude Code's plugin docs for how marketplaces work.

Codex. The repo's install guide lists npx skills add JuliusBrussee/caveman -a codex -g for skills only. Codex calls skills with $, so you type $caveman. See Codex skills and, for Cursor, Cursor skills.

The full installer, npx @caveman-ai/cli@latest, sets up the whole toolkit across every agent it finds and needs Node.js 22.13 or later. That is a bigger change than adding one skill.

How do you use it?

Type /caveman ($caveman in Codex) if it doesn't start on its own, then work as usual. Say stop caveman to go back.

Example prompts (suggestions, not quotes from the repo):

• /caveman
• /caveman Why is this Jest test timing out only in CI?
• /caveman status
• Review this diff for bugs. Keep it caveman.
• stop caveman

Questions that need explaining still work. The rules say clarity wins over compression, and a confused user gets full sentences until the point lands.

In Claude Code, if you installed the plugin but want sessions to start in normal prose, set {"defaultMode":"manual"} in .caveman.json for one project or ~/.config/caveman/config.json for your user, per the skill's README. Then /caveman turns it on when you want it.

Tips and limits

The repo publishes an unusually frank Honest Numbers page. From it and the README, not from testing:

• Small gain over "be concise". Measured once on claude-opus-5-5 with ten prompts in a single run, caveman produced 3% fewer output tokens at the median than a plain Answer concisely. instruction, "inside the noise." ultracave cut 35% and megacave 9%. The numbers come from the 3.1.0 rules and had not been re-run on later versions.
• It costs input. The README says the rules cost about 1,160 tokens of input, though the Honest Numbers page lists the billed input cost as not measured yet. On short Q&A that overhead can outweigh what it saves; the page cites a user who measured a net loss.
• Per-request billing gains nothing. In the README's words, "if you pay per request (GitHub Copilot), shorter answers save nothing."
• It only shrinks output. The skill does not compress input, files or thinking tokens. The repo's separate proxy targets those.
• No quality claim. "No quality-equivalence claim is published." The page's rule of thumb: compare provider-billed totals with and without caveman on the same task, and turn it off if it costs more.

Readability is the other reason to use it: shorter replies are faster to read, whatever they do to the bill.

License. Everything in the repo is Apache-2.0 from Caveman 3.0.0 on, per LICENSING.md; earlier releases keep the terms they shipped with. "Caveman" is a trademark of Julius Brussee and the license grants no trademark rights: naming a fork in a way that implies Caveman sponsorship needs written permission.

Save it in your Kipt library

In Kipt, open Skills and paste the GitHub folder link into the import box. Kipt reads the SKILL.md and README. The repo holds many skills, so Kipt shows a list and you can take ultracave and megacave with it. Your copies live in your own private GitHub repo and install into Claude Code or Codex with one command:

npx skills add https://github.com/<you>/<repo>.git --all -g

Kipt is not affiliated with Julius Brussee or the Caveman project.

Similar skills

• Claude Code output styles: Claude Code's built-in way to change how replies read, without a skill.
• Humanizer skill: rewrites text to sound less machine-made, for prose that leaves the chat.
• Writing guidelines skill: plain, direct writing rules for documents.
• Claude Code context window: where tokens really go in a long session.