Development & Code COMMUNITY New

Caveman

Caveman shortens the output of Claude Code and its subagents, writes terse commit messages and reviews and condenses memory files so your context lasts longer.

Level ●● Intermediate
Saves bloated agent output in your main context
Updated 2026-11-07
Available in: claude-code

What is Caveman?

Caveman is a skill pack for Claude Code that cuts the output of Claude and its subagents down sharply, so the context of a session holds out longer. The pack combines several skills: short commit messages following Conventional Commits, code reviews with one line per finding, condensing memory files such as CLAUDE.md and subagents that hand back results as compact tables with path and line number.

Another part of the pack works with Caveman Cloud. These skills send a repository's LLM requests via a gateway for measurement, assign calls to individual workflows and analyse cost, latency, errors and savings. Whenever a skill wants to change code, it first shows you a proposal and only applies it once you agree.

Caveman suits developers who run long Claude Code sessions, hit context limits or want to lower their token costs. To condense files you need Python 3, because the skill calls a bundled script. The skills for measurement, workflows and experiments require Caveman Cloud.

Who it is for

Developers and teams who work heavily with Claude Code and want to reduce and measure context usage or LLM costs in their repositories.

Use cases

  • You are working on a bug fix in a long Claude Code session and want to hand search, edit and review to subagents without filling up your context.
  • Your CLAUDE.md has grown and loads in every session. You want to condense it so it costs fewer input tokens.
  • Your team runs several LLM workflows and wants to see which workflow causes which cost, latency and errors.

What Caveman does

cavecrew: three compact subagents

An investigator locates code and returns a table with path and line, a builder edits one or two files, a reviewer checks diffs. Their results take up far less room in your main context.

caveman-commit for commit messages

Writes commit messages in Conventional Commits format with the focus on why a change was made. The skill only produces the text, it does not stage files or run a commit.

caveman-review for terse code reviews

With '/caveman-review' you get each finding on a single line naming the location, the problem and a suggested fix, ordered by file and line.

caveman-compress for memory files

With '/caveman-compress' you condense files such as CLAUDE.md or todo lists. Code blocks, paths, commands and URLs stay untouched, and a readable backup is stored outside the project.

Caveman Cloud: cost per workflow

caveman-setup routes LLM requests via the gateway for measurement, caveman-discover names the workflows in your code, caveman-evidence-review reads out cost, latency and errors.

caveman-help as quick reference

With '/caveman-help' or the phrase "caveman help", Claude shows you an overview of the Caveman skills and their commands.

Say this to activate the skill

"Write a commit message for my changes""Run /caveman-review on this diff""Compress my CLAUDE.md with /caveman-compress""Set up caveman for this repository"

Install

npx skills add juliusbrussee/caveman

The command installs every skill in the pack from the author's repository (Vercel skills CLI, needs Node.js), add -g for all projects. We do not host a copy: updates come straight from the author. Need help with setup? How to install skills →

Figures with source and date

GitHub stars
≈ 118,000
Installs, most-installed of 22 skills (caveman)
614,610
On skills.sh since
8 Apr 2026
Last change in the repository
6 Nov 2026
Licence
Apache-2.0

As of 7 Nov 2026. Sources: GitHub, skills.sh. The figures change daily; we show the state of our last lookup.

Checked automatically

Revision of the repository, checked on 7 Nov 2026.

  • Automated check on 2026-11-07 (revision c310c9d): 53 files in the skill folder read by pattern search. It found no hidden characters, no instructions to override rules, no download-and-run and no access to credentials.

Independent checks on skills.sh, as of 7 Nov 2026

Gen Agent Trust Hub Pass
Socket Pass
Snyk Warn

Full results on skills.sh ↗

Good to know

  • caveman-compress overwrites the original file. The skill keeps the readable backup in a separate data folder outside your project.
  • caveman-compress sends the complete file text to the configured model provider. Only condense files that this provider is allowed to see.
  • Several skills such as caveman-setup, caveman-discover and caveman-manage require Caveman Cloud. The source states no prices for it.
  • The output of cavecrew is very terse. If people are meant to read it directly, rewrite it in full sentences first.
  • One auditor on skills.sh reports a warning (Snyk). Read its notes before bringing the skill into projects with sensitive data.
  • The check is an automated pattern search, not a security guarantee, and reads the skill folder only. The command always installs the author's current version; later versions have not been reviewed by us.

Frequently asked questions about Caveman

What is Caveman?

Caveman is a skill pack for Claude Code that condenses output heavily. It contains skills for short commit messages, compact code reviews, shortening memory files, frugal subagents and a connection to Caveman Cloud for measuring LLM costs.

How do I install Caveman?

You install Caveman with the command 'npx skills add juliusbrussee/caveman'. The command installs all skills of the pack at once, including cavecrew, caveman-commit, caveman-review and caveman-compress.

Is Caveman free?

Caveman is released under the Apache-2.0 licence, so you can use the skills freely. The skills built around Caveman Cloud need that service, and the source gives no details on its terms. Token usage with your model provider is billed as usual.

What data does Caveman send?

Caveman's caveman-compress passes the full text of the chosen file to whichever model provider its script is set up for, which can be the Claude command line tool, Anthropic's API or a service compatible with OpenAI. Before sending, it refuses files whose names suggest secrets and files containing credentials such as private keys or API tokens. The cloud skills route LLM requests via the Caveman Cloud gateway.

How does Caveman differ from the default subagents in Claude Code?

Caveman's cavecrew handles the same jobs as Explore or a reviewer but returns results as short, structured lines. If you want explanations, architecture notes or alternatives, the skill itself recommends using the default agents instead.

Share this skill

About this skill

Caveman comes from juliusbrussee and is an open-source project maintained by its author and community.

Author: juliusbrussee / github.com. Source: github.com/juliusbrussee/caveman. Licence: Apache-2.0.

Curated by Collective Brain, an AI and automation digital agency from Hamburg, together with WhiteFox Automations. More vetted skills in the skills catalog.

Working in Germany, Austria or Switzerland? We help teams roll out Claude, in English or German.

Want this skill tailored to your team?

Caveman is a strong starting point. A skill gets really powerful once it knows your workflows, templates and tone of voice. We build custom skills for teams, from analysis to rollout, and connect them to our work in AI consulting.

Request a custom skill →