What is Caveman Mode in Claude Code?

Caveman Mode (also called Kevin Mode) is a short CLAUDE.md / system-prompt snippet that strips filler, bans preambles and summaries, removes rule-of-three lists, and forces terse output. It reduces output-token volume by 61-75% with no loss of code quality.

How much does Caveman Mode save?

At daily-user Sonnet rates, users report roughly $100-140 per month in savings from Caveman Mode alone, because output tokens are the expensive side of the bill and the mode removes the padding around the actual code.

Does Caveman Mode hurt code quality?

No. It removes the preamble, summaries, and filler around the answer, not the answer itself. You still get the code; you stop paying for the words wrapped around it.

Does Caveman Mode work with OpenClaw?

Yes. OpenClaw reads CLAUDE.md the same way Claude Code does, so the response-rules block tightens OpenClaw agent output too. Put it in the project root and restart the gateway with openclaw gateway restart.

← All guides

Guide July 6, 2026

Caveman Mode for Claude Code: Cut Output Tokens 61-75% (2026)

Caveman Mode (also called Kevin Mode) is a short CLAUDE.md snippet that forces terse output. Users report 61-75% output-token reductions with no drop in code quality.

Bleeding money on Claude Code or OpenClaw?

We apply the full cost checklist live and set up monitoring. See training and cost-audit options →

The fix: drop a response-rules block into your CLAUDE.md that bans preambles, summaries, rule-of-three lists, and filler adverbs. Output-token volume drops 61-75%, which at daily-user Sonnet rates is typically $100-140/month saved — because you were paying for words you never needed. It does not lower code quality; it removes the padding around the code.

The CLAUDE.md snippet

Add this to the CLAUDE.md at your project root (OpenClaw reads it the same way Claude Code does):

# Response rules

- No preamble. No "I'll help you with that."
- No summaries of what you did unless asked.
- No rule-of-three. Give the one best answer.
- No filler adverbs (basically, actually, essentially).
- Code first. Explanation only if requested.
- If the task has a single answer, give only that answer.

That is the whole thing. The viral juliusbrussee/caveman repo popularized it; Decrypt and PCWorld both covered it in April 2026.

Why it saves so much

Most of what a chatty model emits around a code change — the “Here’s what I’ll do,” the recap, the three-bullet summary, the hedging adverbs — is output tokens you pay for. Output tokens are the expensive side of the bill. Caveman Mode cuts the padding, not the substance, so you keep the code and drop 61-75% of the token volume around it.

Stack it with ENABLE_TOOL_SEARCH (which cuts the input side, the tool-schema tax) and you have recovered most of what a fresh session bills.

Does it work with OpenClaw?

Yes. OpenClaw inherits the CLAUDE.md convention, so the same response-rules block that makes Claude Code terse also tightens OpenClaw agent output. Put it in the project root where the OpenClaw gateway runs and restart the gateway (openclaw gateway restart) so the new system prompt is picked up.

OpenClaw Costs Guide (all 10 fixes) — the full cost-optimization hub
ENABLE_TOOL_SEARCH: cut the 45K-token tool tax — the input-side fix
Restart Claude Code sessions at 200K tokens — the behavior fix

Need help?

If you want the whole cost checklist applied for you — Caveman Mode, tool search, session discipline, model routing — we do it live in one session. See training and cost-audit options →

Need OpenClaw fixed live?

Remote rescue sessions for gateway, auth, tunnel, VPS, and model access problems.

See Rescue Session