cavernaman

why use many token when few do trick

Fork of caveman by Julius Brussee (MIT). See NOTICE.

Before/After • Install • What You Get • Benchmarks • Full install guide

---

A Claude Code skill/plugin (also Codex, Gemini, Cursor, Windsurf, Cline, Copilot, 30+ more) that makes agent talk like cavernaman — cuts ~75% of output tokens, keeps full technical accuracy. Brain still big. Mouth small.

Before / After

🗣️ Normal Claude (69 tokens) "The reason your React component is re-rendering is likely because you're creating a new object reference on each render cycle. When you pass an inline object as a prop, React's shallow comparison sees it as a different object every time, which triggers a re-render. I'd recommend using useMemo to memoize the object."	Cavernaman Claude (19 tokens) "New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo."
🗣️ Normal Claude "Sure! I'd be happy to help you with that. The issue you're experiencing is most likely caused by your authentication middleware not properly validating the token expiry. Let me take a look and suggest a fix."	Cavernaman Claude "Bug in auth middleware. Token expiry check use < not <=. Fix:"

Same fix. 75% less word. Brain still big.

┌─────────────────────────────────────┐
│  TOKENS SAVED          ████████ 75% │
│  TECHNICAL ACCURACY    ████████ 100%│
│  SPEED INCREASE        ████████ ~3x │
│  VIBES                 ████████ OOG │
└─────────────────────────────────────┘

Pick your level of grunt — lite (drop filler), full (default cavernaman), ultra (telegraphic), or wenyan (classical Chinese, even shorter). One command switch. Cost go down forever.

Like this trick? Now get whole agent — caveman-code This skill shrink what agent say. caveman-code shrink everything — full terminal coding agent, caveman top to bottom. ~2× fewer tokens than Codex on identical tasks. 20+ providers · plan mode · autopilot goal loop · MIT. npm install -g @juliusbrussee/caveman-code ▶ Try caveman-code now → — why use many token when whole agent save

Install

One line. Find every agent. Install for each.

# macOS / Linux / WSL / Git Bash
curl -fsSL https://raw.githubusercontent.com/jagoff/cavernaman/main/install.sh | bash

# Windows (PowerShell 5.1+)
irm https://raw.githubusercontent.com/jagoff/cavernaman/main/install.ps1 | iex

~30 seconds. Needs Node ≥18. Skip agent you no have. Safe to re-run.

Trigger: type /cavernaman or say "talk like cavernaman". Stop with "normal mode".

One agent only, manual command, or any of 30+ other agents → INSTALL.md. Install break? Open agent, say "Read CLAUDE.md and INSTALL.md, install cavernaman for me." Agent fix own brain.

What You Get

Skill	What
/cavernaman [lite|full|ultra|wenyan]	Compress every reply. Levels stick until session end.
/cavernaman-commit	Conventional Commit messages, ≤50 char subject. Why over what.
/cavernaman-review	One-line PR comments: L42: 🔴 bug: user null. Add guard.
/cavernaman-stats	Real session token usage + lifetime savings + USD. Tweetable line via --share.
/cavernaman-compress <file>	Rewrite memory file (e.g. CLAUDE.md) into caveman-speak. Cuts ~46% input tokens every session. Code/URLs/paths byte-preserved.
caveman-shrink	MCP middleware. Wraps any MCP server, compresses tool descriptions. npm.
cavecrew-*	Cavernaman subagents (investigator/builder/reviewer). ~60% fewer tokens than vanilla, main context lasts longer.

Statusline badge — Claude Code shows [CAVERNAMAN] ⛏ 12.4k (lifetime tokens saved). Updates every /cavernaman-stats run. Set CAVEMAN_STATUSLINE_SAVINGS=0 to silence.

Auto-activate every session: Claude Code, Codex, Gemini (built-in). Cursor / Windsurf / Cline / Copilot get always-on rule files via --with-init. Other agents trigger with /cavernaman per session. Full feature matrix in INSTALL.md.

Benchmarks

Real token counts from the Claude API. Average 65% output reduction across 10 prompts (range 22-87%).

Task	Normal	Cavernaman	Saved
Explain React re-render bug	1180	159	87%
Fix auth middleware token expiry	704	121	83%
Set up PostgreSQL connection pool	2347	380	84%
Explain git rebase vs merge	702	292	58%
Refactor callback to async/await	387	301	22%
Architecture: microservices vs monolith	446	310	30%
Review PR for security issues	678	398	41%
Docker multi-stage build	1042	290	72%
Debug PostgreSQL race condition	1200	232	81%
Implement React error boundary	3454	456	87%
Average	1214	294	65%

Raw data and reproduction script: benchmarks/. Three-arm eval harness (baseline / terse / skill) lives in evals/ — cavernaman compared against Answer concisely. not against verbose default, so the delta is honest.

caveman-compress receipts (real memory files):

File	Original	Compressed	Saved
claude-md-preferences.md	706	285	59.6%
project-notes.md	1145	535	53.3%
claude-md-project.md	1122	636	43.3%
todo-list.md	627	388	38.1%
mixed-with-code.md	888	560	36.9%
Average	898	481	46%

Important

Cavernaman only affects output tokens — thinking/reasoning tokens untouched. Cavernaman no make brain smaller. Cavernaman make mouth smaller. Biggest win is readability and speed, cost savings a bonus.

A March 2026 paper "Brevity Constraints Reverse Performance Hierarchies in Language Models" found that constraining large models to brief responses improved accuracy by 26 points on certain benchmarks. Verbose not always better. Sometimes less word = more correct.

How cavernaman differ from caveman

cavernaman a fork of caveman by Julius Brussee (MIT). caveman invent whole thing — the voice, the levels, the install matrix, the hook that make Claude talk terse from message one. Big thanks. cavernaman stand on caveman shoulder. Go star caveman. 🙏

cavernaman add a "token-savings v2" pass on top — every number a real run:

• ultra v2 — response-shape templates, markdown-strip, symbol swaps, single-word answers. +23% tighter output vs caveman shipped ultra (median +19%).
• caveman-shrink v2 — compress prose in tool-call results that used to pass through untouched. +14% on prose; structured data (JSON/tables/code) stay byte-identical.
• Injected-ruleset trim — per-session ruleset cost drop 2–8% per level, no behavior change.
• Three-axis measurement harness — savings checkable, floor cases shown not hidden.

cavernaman cut tokens on three independent levers, each measured separately:

Full breakdown, repro commands, honest caveats → docs/cavernaman-vs-caveman.md.

How It Work

1. Install drop skill file in agent.
2. Skill tell agent: drop filler, keep substance, use fragments.
3. For Claude Code, hook also write tiny flag file each session — agent see flag, talk cavernaman from message one. No need say /cavernaman.
4. Stats command read Claude Code session log, count tokens saved, write number to statusline.
5. Cavernaman-compress sub-skill rewrite memory files (CLAUDE.md, project notes) so each session start with smaller context. Save tokens forever, not just one reply.

Maintainer detail (hook architecture, file ownership, CI sync) live in CLAUDE.md.

Lobster, Meet Rock 🦞

OpenClaw the self-host gateway. One box, many agent inside (Claude Code, Codex, Pi, OpenCode), wired to your Slack / Discord / iMessage / Telegram / whatever. Tagline: "The lobster way." Lobster strong. Lobster smart. Lobster also talk a lot.

Cavernaman teach lobster brevity — same canonical installer, scoped to one agent:

# macOS / Linux / WSL
curl -fsSL https://raw.githubusercontent.com/jagoff/cavernaman/main/install.sh | bash -s -- --only openclaw

# Windows (PowerShell): no Node? install Node ≥18 first, then
npx -y github:jagoff/cavernaman -- --only openclaw

Two thing happen, no more:

1. Skill drop at ~/.openclaw/workspace/skills/caveman/SKILL.md — spec-correct frontmatter (version, always: true), discoverable by openclaw skills list. Skill not auto-inject (OpenClaw load skill on demand) — that why we also do step 2.
2. SOUL.md nudge. Tiny marker-fenced block appended to ~/.openclaw/workspace/SOUL.md. OpenClaw inject SOUL.md into every turn under "Project Context" (12K-per-file, 60K total — block well under). Lobster terse from message one. No /cavernaman per session. No nag.

~/.openclaw/workspace/
├── skills/caveman/SKILL.md   ← full ruleset, on-demand load
└── SOUL.md                    ← <!-- caveman-begin --> ... <!-- caveman-end -->
                                  ↑ auto-inject every turn

Custom workspace path? OPENCLAW_WORKSPACE=/your/path before the command. Uninstall: same one-liner with --uninstall — skill folder gone, SOUL.md block ripped out cleanly, your other workspace content stay untouched. Idempotent re-runs (frontmatter not double-prepended, marker block not duplicated).

Lobster claw still sharp. Lobster mouth now small. Brain still big.

Cavernaman Ecosystem

Five tools. One philosophy: agent do more with less.

Repo	What
cavernaman (you here)	Output compression — why use many token when few do trick
caveman-code	Whole terminal coding agent — why use many token when whole agent can save
cavemem	Cross-agent memory — why agent forget when agent can remember
cavekit	Spec-driven build loop — why agent guess when agent can know
cavegemma	Gemma 4 31B fine-tuned on cavernaman pairs — why prompt every turn when weight remember

Compose: cavekit drive build, cavernaman compress what agent say, cavemem compress what agent remember, cavegemma bake compression into weight, caveman-code ship it all as one terminal agent. One rock. Two rock. Three rock. Four rock. Five rock. That it.

Links

• INSTALL.md — full install matrix, all flags, per-agent detail
• CONTRIBUTING.md — how to send patch
• CLAUDE.md — maintainer guide (file ownership, hook architecture, CI)
• docs/ — extra guides (Windows install, etc.)
• Issues — bug, feature, weird behavior

Star This Repo

Cavernaman save you token, save you money. Star cost zero. Fair trade. ⭐

Also by Julius Brussee

• Revu — local-first macOS study app with FSRS spaced repetition. revu.cards

License

MIT — free like mass mammoth on open plain.