Caveman Mode

caveman-claude 🪨

Caveman mode for Claude Code: same technical quality, ~60–65% fewer output tokens in full mode (default). Smart caveman speak prose, code stay normal.

Ported from JuliusBrussee/caveman, enhanced with BANNED patterns + pre-response filter for reliable compression.

User: Why is my React component re-rendering?

Caveman: New object ref each render. Inline object prop = new ref = re-render. Use useMemo.

What you get

Piece	Path	What it does
Output style	output-styles/caveman.md	Persistent caveman mode via /output-style Caveman (recommended)
Skill	skills/caveman-compress/SKILL.md	Full compression rules; Claude auto-invokes when you ask for brevity
Command	commands/caveman.md	/caveman lite|full|ultra|off toggle in the current session
Agent	agents/caveman.md	Subagent that does research/implementation and reports in caveman

Install

As a plugin (recommended — inside Claude Code):

/plugin marketplace add liusc45/caveman-claude
/plugin install caveman@caveman-claude

Or manual copy into ~/.claude:

git clone https://github.com/liusc45/caveman-claude.git /tmp/caveman-claude && bash /tmp/caveman-claude/install.sh

Use

/output-style Caveman    # persistent for the project (recommended)
/caveman ultra           # maximum grunt, this session
/caveman lite            # professional, no fluff
/caveman off             # normal mode

You can also just ask: "caveman mode" or "less tokens" — the skill triggers on those.

Intensities

Level	Target	What changes
lite	~25%	No filler/hedging. Keep articles + full sentences. Professional but tight
full (default)	~60–65%	Drop articles, fragments OK, short synonyms. Answer first, 1-line bullets, zero padding
ultra	~80%	All full rules + prose abbreviations, strip conjunctions, arrows X → Y, one word when enough

Boundaries

Caveman applies to prose only. Code, git commits, PR descriptions and file contents stay normal. Error messages quoted exact. Auto-clarity: normal prose for security warnings, architecture decisions, onboarding.

Why output style + skill + command?

• Output style is Claude Code's native mechanism for tone — it edits the system prompt while keeping all coding instructions (keep-coding-instructions: true), so it persists across the session and survives compaction.
• Skill lets Claude self-activate when you ask for brevity in plain words.
• Command gives you per-session intensity control without changing your default style.

See also

• caveman-opencode — same thing for opencode
• caveman-code — the original standalone CLI (adds LLMLingua context compression and tool-output compression on top of the speech style)

License

MIT. Compression rules from caveman-code by Julius Brussee (MIT).