The caveman Skill for Claude Code: Cut Output Tokens Without Losing Substance

Long agent sessions die by a thousand polite paragraphs. The caveman skill attacks exactly that: it strips filler, hedging, and pleasantries from every response while leaving the technical payload untouched.

Key takeaways

• caveman by juliusbrussee has 319,221 installs and 75,496 GitHub stars, with 92 new installs in the last day (skills.sh registry, GitHub, 2026-07-06).
• Its SKILL.md targets roughly 75 percent output-token reduction by dropping articles, filler, and hedging while keeping code blocks, commands, and error strings byte-for-byte exact.
• It ships intensity levels - lite, full (the default), and ultra - plus wenyan variants for even tighter compression.
• Once triggered it stays active every response until you say "stop caveman" or "normal mode".
• The repo is actively maintained: last push 2026-06-12 (GitHub), and the skills.sh audit trail rates it LOW risk.

What is the caveman skill?

caveman is a communication-style skill: it rewrites how the agent talks, not what it does. Prose gets compressed to fragments ("why use many token when few token do trick"), but anything technical - code, shell commands, file paths, error output - is preserved exactly. That distinction is what makes it usable for real work rather than a novelty.

How do you install and trigger caveman?

Install from the project root:

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

Trigger phrases include "caveman mode", "talk like caveman", "less tokens", or "be brief". On Claude Code a hook writes a small flag file each session, so the agent can talk caveman from the first message without an explicit command. The official Claude Code skills docs explain where installed skills live and load.

How much does caveman actually save?

The skill's own target is roughly 75 percent fewer output tokens (SKILL.md). Real savings vary with how verbose the baseline reply would have been - short factual answers compress less than long explanations. Two caveats worth knowing before you rely on it for cost control: it only shrinks output tokens, not input or reasoning tokens, and the skill definition itself adds a small amount of input per turn. In long sessions where the context window is the constraint, the compression usually pays for itself many times over.

What are the intensity levels?

Three main levels, in increasing aggressiveness: lite keeps full sentences but cuts filler and hedging; full (the default) drops articles and allows fragments; ultra abbreviates aggressively and uses arrows for causality. The wenyan-lite, wenyan-full, and wenyan-ultra variants push further using classical-Chinese-style compression. Start with full; switch to lite if you find yourself re-reading answers.

When should you not use caveman?

Skip it when the output is the deliverable: documentation, commit messages, PR descriptions, user-facing copy, or anything a teammate will read verbatim. It also fights skills whose value is structured, readable prose. For pipeline-facing work, browse the dev-tools shortlist instead - compression belongs in conversations, not artifacts.

Common pitfalls

• Leaving caveman on while generating docs or commit messages, then shipping fragments to teammates.
• Expecting total-cost cuts equal to the output reduction - input and reasoning tokens are untouched.
• Stacking it with other style skills; the last one loaded tends to win and the result is unpredictable.

Few token. Same substance. That is the whole trick, and 319,221 installs say it works.

Common questions