Traction scoreGitHub stars can be faked, so popularity alone can be misleading. Traction Score looks for broader signs of recent attention, adoption, and active maintenance.
Caveman provides a proxy and prompt engineering framework that minimizes the number of tokens used during LLM communication. By adopting a compressed, primitive style of interaction, it helps AI agent builders and prompt engineers maintain utility while significantly lowering operational costs.
The system focuses on removing unnecessary words from agent prompts and responses, making the exchange of information between coding agents more efficient without sacrificing the core intent of the instruction.
Compatible AgentsThe repository documents support for these agents. The skills may also work with other agents that can load SKILL.md files, but they may need some setup or small changes.
Follow the documented setup, then try a first task.
I’m using
Unified installer for macOS, Linux, WSL, or Git Bash
Installs: Complete Caveman installation for every supported agent detected on the machine; the installer uses user scope except for Replit, whose documented location is the current project's .agents/skills. · Claude Code, Cursor, Codex, other supported installed agents
Before you start
Node.js 22.13 or newer
macOS, Linux, WSL, or Git Bash
Platform: macOS / Linux / WSL / Git Bash
Run this command in a terminal. If installing for Replit, run it from that project's Shell because Replit uses the project-local .agents/skills location.
Installs: Complete Caveman installation for every supported agent detected on the machine; the installer uses user scope except for Replit, whose documented location is the current project's .agents/skills. · Claude Code, Cursor, Codex, other supported installed agents
Before you start
Node.js 22.13 or newer
Windows PowerShell 5.1 or newer
Platform: Windows
Run this command in PowerShell 5.1 or newer. If installing for Replit, run it from that project's Shell because Replit uses the project-local .agents/skills location.
Installs: Complete caveman plugin from the Caveman Claude Code marketplace; the listed skillPaths are a bounded starter selection from the plugin source. · Claude Code
Before you start
Claude Code with the claude command available
Run this complete command in a terminal to add the author's marketplace and install the caveman plugin.
claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman
Installs: Complete Caveman repository skill bundle in the Codex user skill directory; the listed skillPaths are a bounded starter selection. · Codex CLI
Before you start
Node.js environment with npx available
Codex CLI installed
Run this command in a terminal. The -g flag selects the agent's user skill directory rather than the current project.
Installs: Complete Caveman repository skill bundle in the Windsurf user skill directory; the listed skillPaths are a bounded starter selection. · Windsurf
Before you start
Node.js environment with npx available
Windsurf installed
Run this command in a terminal. The -g flag selects the agent's user skill directory rather than the current project.
npx skills add JuliusBrussee/caveman -a windsurf -g
Use the investigate-first skill to diagnose [observed symptom] in [repository or path]. Use these reproduction steps and logs: [steps and evidence]. Separate observations from inferred causes, trace the relevant state and ownership boundaries, rank hypotheses by evidence and cheap falsification value, and report the most credible cause and proof. Do not edit or implement a fix.
Use the caveman-commit skill to write a Conventional Commits message for this diff: [paste diff]. Apply this project's commit convention if provided: [convention]. Capture intent and any non-obvious why; include a body only when the skill's rules require it. Output the message ready to paste, but do not stage files or run git commit.
Relevant behavioral expectations or acceptance criteria
Use the caveman-review skill to review [paste or identify the diff]. Return one actionable line per finding with the exact file and line, the problem, and a concrete fix. Use severity prefixes when findings are mixed. Do not edit the code, run linters, approve, or request changes.