Caveman: A Claude Code Skill That Saves AI Agents About 65% of Output Tokens with Caveman-Style Minimal Expression

Caveman is a Claude Code skill that compresses output tokens using "caveman-style minimal expression." Its best-fit scenario is clear: you use AI agents heavily for coding or daily tasks, your bill is dominated by the model's reply text, and you want to cut output tokens without switching models (officially claimed at about 65%). But there's a key caveat you must know before using it: it only compresses output tokens — input and reasoning tokens are completely unaffected — and the skill itself adds roughly 1–1.5k input tokens per turn. If your costs come mainly from long-context input, it won't help you.

Who It's For

• Good fit: developers who use Claude Code and similar agents at high frequency daily, want to reduce reply-text costs, and can accept a more terse (even telegraphic) output style.
• Good fit: users willing to first try the intensity levels (lite/full/ultra, plus the wenyan series) on low-risk tasks before deciding whether to keep it enabled long-term.
• Not a fit, or use with caution: scenarios where costs come mainly from input tokens (long context, large files) — the official docs explicitly state input and reasoning tokens are unaffected, and the skill adds input per turn.
• Not a fit: tasks requiring the model's output to remain complete, formal, and rich in detail — minimal style is itself its compression mechanism.
• Unknown: compression effectiveness for specific codebases, languages, or task types; the evidence provides no per-scenario data, so test it yourself.

First Try in 3 Minutes

The recommended main path is installing the skill itself via the skills CLI (not the proxy, and no need to change your global environment):

1. Prerequisite: an agent that supports the skill is installed (officially listed: Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, and 30+ other agents).
2. Install location: run the skills CLI install command in your development environment:

npx skills add JuliusBrussee/caveman

1. Trigger: open your agent; if the skill doesn't activate automatically, type /caveman to enable it manually. For the first run, it's recommended to use the low-intensity level /caveman lite to minimize the impact on output style.
2. Low-risk first task: have the agent do a read-only, reversible small task, e.g., "explain what this code does in the current mode." Success signal: the agent's reply clearly becomes terse, telegraphic style (or classical-Chinese style under the wenyan level), with no errors.
3. Adjust and exit: use /caveman full or /caveman ultra to increase intensity; use /caveman off (or say "normal mode") to restore normal output.

On "token savings within a session": the official page's core promise for the skill is compressing output tokens by about 65%, but the evidence doesn't state verbatim how the specific savings figure is displayed or whether statusline or hooks are required — observe it in your actual sessions.

The official docs also offer two other installation methods, noted only as secondary options: one is the proxy product npm install -g @caveman-ai/cli && caveman setup --install, after which you launch the agent with caveman claude (also supports codex, gemini, aider, opencode, hermes, openclaw); the other is the full install script (requires Node.js 18+, integrates with Claude Code hooks and statusline, and is re-runnable). Piping a script executes remote code on your machine and modifies agent configuration — a heavier action than a skill install — so don't start your first attempt there.

Use Cases

Case 1: Reducing everyday Q&A costs. You often ask Claude Code technical questions, and the replies are long and token-hungry. After installing, type /caveman lite, then ask a routine question. It will run the skill's output-compression workflow and rewrite the reply in minimal style; the expected result is the same answer points with noticeably fewer output tokens. If it feels too sparse, switch to /caveman full; to restore the original, use /caveman off.

Case 2: Controlling costs in long coding sessions. You plan to have the agent run a round of refactoring or a batch of small changes. First confirm activation (/caveman or auto-activation), pick an intensity level, then start the task. Throughout the session, the agent's explanations and commentary will be compressed and output tokens will drop; the correctness of the code itself still requires your normal review — the evidence doesn't state that the skill performs additional code verification, so don't assume it does.

Note: the cases above are based on the capabilities described in the official README; how much can be confirmed after step 2 in each case (e.g., exactly how the savings percentage is displayed) is not given in the evidence — rely on actual behavior in your agent.

Usage Boundaries

• Input/reasoning tokens: the official docs explicitly state the skill only compresses output and adds roughly 1–1.5k input tokens per turn; short sessions or input-heavy scenarios may not be worth it.
• File and command access: as a skill, its mode of operation is influencing the agent's output style; whether it directly reads/writes your files or executes commands is not stated in the evidence and is unknown. For a first attempt, choose a read-only task.
• Network access: npx pulls the package from npm during installation; whether the skill itself accesses the network at runtime is not stated in the evidence.
• Installation differences: the skills CLI install is the universal entry point; the proxy version (caveman claude and similar commands) and the full install script are two other official paths with different behavior and impact — don't treat them as the main path or mix them in.
• Clients and toggles: the official docs list Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, and 30+ other agents as supported, controlled via the /caveman command series; whether it works in other environments depends on your own testing.
• Unconfirmed: the 65% savings figure is an official claim, not an independent measurement; the language coverage beyond the Chinese/classical-Chinese levels (wenyan series) and specific differences across agents are not given in the evidence.