~/codigoagente
AI agents

Caveman: the free skill that makes your agent grunt and spend 33% fewer tokens (real test)

At the end of the article you'll find the install command, a ready-to-paste prompt and a .md to download.

Caveman is a skill (and plugin) for Claude Code, Codex, Gemini CLI and thirty-plus other agents that forces your model to answer like a caveman: short sentences, no "sure, happy to help", no decorative commentary and not a single character changed in the code. The numbers are measured in the JuliusBrussee/caveman repo: 35% fewer output tokens with /ultracave, 33.2% fewer input tokens per session thanks to the tool-output compressor and 0.46× the cost of Opus at xhigh when you let Auto pick for you. Cited by Adobe Research, A/B tested by JetBrains and released under a dual Apache-2.0 + MIT license.

Same fix, 63 tokens vs 20. The skill cuts the filler, not the code · codigoagente

What Caveman is and who builds it

  • Repo: github.com/JuliusBrussee/caveman, published by Julius Brussee. Analyzed version: v4.1.0.
  • License: Apache-2.0 by default (LICENSE) with a parallel LICENSE-MIT file, so you can use the skill as MIT too. Commercial projects are covered either way.
  • Install shapes: three, depending on how far you want to go. The skill alone (just the voice), the Claude Code plugin (the voice plus /caveman-review, /caveman-commit, /caveman-stats) or the full installer, which also adds the tool-output compressor, the MCP tools and the cache-warming hooks.
  • What it isn't: Caveman never rewrites your prompt and never edits the generated code. The README cites Adobe's CAVEWOMAN paper: compressing the user's prompt makes answers worse, so the skill only touches what the model *writes* and what it *reads*.
  • Agents covered: Claude Code (native plugin), Codex CLI (SessionStart hook), Gemini CLI (extension), Cursor, OpenCode, Oh My Pi, OpenClaw, Hermes, Antigravity and a long list in INSTALL.md.

The six "rocks" in Caveman 4.1

Since 4.0 the project is split into six pieces that all turn on at install. Each one has its own switch (caveman on <rock>, caveman off <rock>).

The six rocks and the numbers the README publishes · codigoagente

1. Say less: the caveman voice

The original skill: it changes how the agent talks. The rules live in skills/caveman/SKILL.md and there are seven, not seven paragraphs of philosophy:

  1. Answer first. Fixed pattern: [thing] [action] [reason]. [next step].
  2. No ceremony. No "sure", no "happy to help", no "hope this works".
  3. No meaning ever dropped. Articles can go. not, never and numbers never do.
  4. Payload stays verbatim. Code, commands, paths and errors untouched.
  5. One idea per sentence. Built on ASD-STE100, the controlled English used in aviation manuals: 20 words max, active voice, one term per thing.
  6. Knows when to stop. Security warnings, irreversible steps and confused users get full sentences back.
  7. Your prompt is yours. Rewriting it backfires, Adobe says.

Two cousins: /ultracave goes even shorter (35% less output than "Answer concisely." over 10 questions) and /megacave answers in classical Chinese (13 tokens per reply). Both are triggered with /caveman ultra or /caveman wenyan.

2. Read less: compress what enters the model

The least obvious piece and the biggest saver. When the agent calls bash, grep, read_file or an MCP tool, the output runs through a local filter before it hits the context: summarised logs, JSON with redundant keys dropped, CSV with ignored columns removed and test output pruned to what fails. On the repo's own bench (6 synthetic large-output tasks, 18 pairs) the agent gets 33.2% fewer input tokens per session without missing an answer.

3. Auto: picks model and effort for you

Decides what model and what reasoning_effort to use, question by question. On 85 coding tasks it measures 0.46× the cost of Opus always at xhigh, with a quality gap of ±7 points that is still unresolved. The only piece that needs an account on caveman.so, and that account is free.

4. Blocks: code mode with reusable scripts

When your agent writes a bash or Python script to solve a task, Caveman keeps it. Next time the agent has to run the same thing (the same test, the same Postgres query) it reruns the saved script instead of generating it again. README example: running a test drops from 397 tokens to 76 (| tail -20 vs the saved block).

5. Waste fixes

Local analytics on where your tokens go, with a caveman stats panel. It also keeps the prompt cache warm when you pause: come back minutes later and the session does not recompact the context from scratch.

6. Browse

Playwright with caveman mode: web pages come back compressed. In one README example a 200-row table goes from 15,704 tokens to 121 when the request carries a filter.

Installing Caveman on Claude Code (and on the rest)

The author recommends the unified installer. On macOS, Linux or WSL with curl:

terminalsh
curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/v4.1.0/install.sh | bash

On Windows with PowerShell 5.1+:

terminalsh
irm https://raw.githubusercontent.com/JuliusBrussee/caveman/v4.1.0/install.ps1 | iex

The script detects each agent you already have installed and uses its native path: claude plugin install for Claude Code, gemini extensions install for Gemini CLI, a SessionStart hook for Codex. If you don't want to curl … | bash, download first and read: every runtime binary is checksummed against a list signed with a key built into the CLI, and the hooks are verified against the same v4.1.0 of the installer.

If you only want the voice (no compressor, no hooks), use the minimal path:

terminalsh
npx skills add JuliusBrussee/caveman -g

For the Claude Code plugin only:

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

After installing, run caveman setup to pick which rocks to keep on. caveman status prints the state and caveman doctor checks dependencies (Node 22.13+, Git Bash on Windows, write access to ~/.caveman).

The skill up close: what changes in your sessions

  • Activation. When you open Claude Code with the plugin it loads skills/caveman/SKILL.md, which is what the model reads. While the skill is active, Claude replies in caveman mode. Turn it off with /caveman stop or by saying "normal mode".
  • Commands added. The plugin ships three slash commands on top of the voice:
  • /caveman-commit writes the commit message in Conventional Commits with the same terse voice.
  • /caveman-review scans the git diff for "fluff" (decorative comments, unused imports, forgotten logs) and lists them flat.
  • /caveman-stats prints how many tokens the session saved.
  • Aliases. /ultracave and /megacave are separate skills with the same hard rules; answering in classical Chinese (wenyan) shaves another 7% in the author's internal runs.
  • cavecrew subagents. If your agent supports native subagents (Claude Code does), the installer registers an investigator, a builder and a reviewer, all talking short. Handy for /plan and long refactors.

What the project measures (and what it doesn't)

The README.md ships the numbers with their conditions, and the "HONEST-NUMBERS" section is required reading before attributing miracles:

  • Terminal-Bench 4, GPT-6.1 Sol vs Codex: 21% cheaper per solved task ($1.21 vs $1.52, 90 of 192 tasks solved by both).
  • Terminal-Bench 4, Opus 5.5 vs Claude Code: tied inside the noise (97/180 vs 83/177, $5.19 vs $6.04 per solved task).
  • Adobe Research, CAVEWOMAN paper: caveman-style output on 8 models and 5 datasets, cost cut between 1.4× and 2.4×, up to 3× on edge cases.
  • Elastic, Elasticsearch Labs: 63.6% fewer response tokens on 8 MCP scenarios, "zero information loss".
  • JetBrains, 86 real coding tasks: no measurable quality drop (p = 0.82), 8.5% fewer output tokens.

Three caveats worth printing on the wall:

  1. The rules cost about 1,160 input tokens every time they load. If you pay per request (GitHub Copilot) instead of per token, shorter replies save you nothing.
  2. "Be concise" is the real baseline. Caveman cuts 3% more at the median, /ultracave cuts 35% more.
  3. New models (Claude 5.5, GPT-6.1) already answer shorter by default; part of the saving overlaps with what the provider already does.

Problems we hit

  • npm 12 breaks the npx -y github:... commands. Any npm 12 or newer stops with npm error code EALLOWGIT. Fix: use the pinned v4.1.0 one-liner, which detects the version and adds --allow-git=root, or pass the flag by hand.
  • Node < 22.13. The compressor and Auto refuse to start. On Linux: nvm install 22 and rerun caveman setup.
  • Bun-only machines. The installer supports Bun, but native hooks need Node ≥ 18; with Bun alone the installer lands the skills and warns about what it skipped.
  • Does the plugin stomp on your CLAUDE.md? No: the installer adds CAVEMAN.md and references it from CLAUDE.md with a single line; delete that line to disable it without uninstalling.
  • Black box for actual billing. caveman stats counts with tiktoken o200k, not the provider's invoiced counter. It gives order of magnitude, not the receipt.

FAQ

Can I ship it in client projects? Yes. Apache-2.0 and MIT allow commercial use and modification with no obligation to publish derivatives. The only thing you can't do is use the Caveman name or logo to endorse your own product.

Do I lose safety warnings? No. Rule 7 forces the model to go back to full sentences on irreversible steps, security warnings and confused users.

Does it work with a Claude Pro login? Yes. Caveman reuses the Claude Pro / Max login your CLI already has. The only piece that needs a separate account is Auto, and that account is free on caveman.so.

I don't like broken English. The grunt is theatre; in practice it answers in short sentences, not grunts. If you want the trim without the theatre, install the plugin and use /ultracave: same cut, more serious voice.

Verdict

Caveman is the first skill that bundles the three things people have been asking for since agents shipped: short output, controlled input and cheap model routing by default. All three turn on with a single curl … | bash, and each rock can be switched off when it gets in the way. It never rewrites your prompt, never touches your code and publishes its numbers including the cases where it loses: for a blog or a SaaS that pays Claude in tokens, the question is why not have it on.

The logical next step is to pair it with a serious review skill like Open Code Review on Claude Code or Superpowers, or combine Caveman with Impeccable so that short output doesn't fight your design.

// resource · use it now

Caveman ready to use in Claude Code

Install the plugin, copy the prompt for your case and paste it as the first message of the session. It works with your Claude Pro or Max plan; no API key needed.

Repository license: Apache-2.0 + MIT

1. Install

Claude Code plugin
claude plugin marketplace add JuliusBrussee/caveman
claude plugin install caveman@caveman
Skill only (any agent)
npx skills add JuliusBrussee/caveman -g

2. Pick your case

Paste this prompt into your agent
We are starting this project with the caveman skill active to spend fewer tokens.

1. Check that the caveman skill is loaded. If it isn't, tell me the command to install it and stop here.
2. Turn on /caveman mode (not ultra) for the whole session: short answers, but code, commands, paths and errors exact and untouched.
3. Switch back to full sentences for security warnings, destructive commands or if I say I don't understand something.
4. Create a minimal CLAUDE.md with: stack, install, dev and test commands, and this rule: "Answer in caveman mode except for security warnings and irreversible steps".
5. Before writing code, give me the plan in 5 lines max and wait for my OK.
6. Use /caveman-commit for every commit (Conventional Commits).

3. How to use it

  1. Open Claude Code in the empty project folder and paste the prompt as the first message.
  2. Review the CLAUDE.md it creates: the caveman rule stays there for future sessions.
  3. If something is hard to follow, type "normal mode" or /caveman stop, and go back to /caveman whenever you like.
  4. When you close the session, run /caveman-stats to see how many tokens it saved.
Prompt and short guide in one fileRepositoryJuliusBrussee/caveman

Help us grow

codigoagente is free, with no paywalls. If this article helped you, any of these keeps the blog going.

// newsletter

One email a week with the repos worth your time

The week's new articles, tested against real code. No spam; unsubscribe in one click.

AI agents

guide · The complete ECC guide for Claude Code · #8

AgentShield, ECC's scanner, checks CLAUDE.md, settings.json, MCP, hooks and agents. We ran it on a deliberately vulnerable setup and on ECC's own install.
  • Security and code review
  • Claude Code
read article
AI agents

guide · The complete ECC guide for Claude Code · #7

We measured ECC's context footprint with /ecc:context-budget --audit (about 12,500 tokens on the minimal profile) and how to trim it: model, compaction and MCP.
  • Claude Code
  • Skills for your agents
read article
AI agents

guide · The complete ECC guide for Claude Code · #6

What ECC's 28 hooks do, how GateGuard stops the first edit, how to save and resume sessions and how continuous learning instincts work, with a real test.
  • Claude Code
  • Skills for your agents
read article