Engineering TokensCompressionProductivity

Caveman

Cut ~75% of output tokens by having the agent talk like caveman — full technical accuracy, zero fluff. Intensity levels: lite, full, ultra, and classical Chinese (文言文).

A community skill by Julius Brussee that makes your AI agent communicate like a caveman — stripping articles, filler, hedging, and pleasantries while preserving every byte of technical substance. Benchmarked at 65% average output token savings (up to 87%). Ships with caveman-commit for terse commit messages, caveman-review for one-line PR comments, and caveman-compress which cuts CLAUDE.md / memory file read tokens by ~46% each session. Works on Claude Code, Cursor, Windsurf, Copilot, Cline, Codex, Gemini CLI, and 40+ more agents.

Added Apr 9, 2026

$npx skills add JuliusBrussee/caveman

Capabilities Real Examples Get Started Models & Context Works Great With Good to Know

What This Skill Can Do

Concrete capabilities you get when you install this skill.

Reduce output tokens ~65% on average (up to 87%) with no loss in technical accuracy

Switch intensity on the fly: lite (drop filler), full (classic caveman), ultra (maximum compression)

文言文 (Wenyan) mode — classical Chinese literary compression for maximum brevity

caveman-commit: generate terse, conventional-commits-style messages ≤50 char subject

caveman-review: one-line PR comments (e.g. L42: 🔴 bug: user null. Add guard.)

caveman-compress: compress CLAUDE.md and memory files ~46% — fewer input tokens every session

Auto-clarity mode for destructive ops, security warnings, and multi-step sequences

Works on Claude Code, Cursor, Windsurf, Copilot, Cline, Codex, Gemini CLI, and 40+ more agents

Real Conversations

See how the Caveman skill responds to actual requests. These are representative examples of what your AI agent will produce.

You

/caveman

Why is my React component re-rendering every time the parent renders, even though the props haven't changed?

Agent

Caveman mode: full. Active every response.

Component re-render each parent render = prop reference change.

Root cause: Object/array/function prop created inline → new reference each render → React shallow compare see different → re-render.

// BAD — new object every render
<Child config={{ timeout: 5000 }} />

// GOOD — stable reference
const config = useMemo(() => ({ timeout: 5000 }), []);
<Child config={config} />

Also check:

useCallback missing on function props
Context value not memoized → all consumers re-render
Redux useSelector returning new object each call → add shallowEqual

Quick diagnosis:

import { useRef } from 'react';

function Child({ config }) {
  const renders = useRef(0);
  console.log('render #' + ++renders.current, config);
  // ...
}

Log shows same value, different reference? → useMemo fix. Value actually changing? → parent has bug.

React DevTools Profiler → "Why did this render?" tab → shows exact prop that changed.

You

/caveman ultra

I'm seeing intermittent data corruption in my PostgreSQL table. Two workers process the same job queue. Sometimes both pick up the same job and write conflicting data.

Agent

Ultra mode: active.

Race: 2 workers SELECT same job → both process → conflict.

Fix: SELECT ... FOR UPDATE SKIP LOCKED

-- Worker query (atomic claim)
UPDATE jobs
SET status = 'processing', worker_id = $1, claimed_at = NOW()
WHERE id = (
  SELECT id FROM jobs
  WHERE status = 'pending'
  ORDER BY created_at
  LIMIT 1
  FOR UPDATE SKIP LOCKED
)
RETURNING *;

SKIP LOCKED = worker 2 skips rows worker 1 locked → no double-claim. Single atomic UPDATE = no separate SELECT+UPDATE gap.

Schema additions:

ALTER TABLE jobs ADD COLUMN worker_id TEXT;
ALTER TABLE jobs ADD COLUMN claimed_at TIMESTAMPTZ;
ALTER TABLE jobs ADD COLUMN status TEXT DEFAULT 'pending'
  CHECK (status IN ('pending','processing','done','failed'));
CREATE INDEX ON jobs(status, created_at) WHERE status = 'pending';

Retry on crash:

-- Run every 30s — reclaim stuck jobs
UPDATE jobs SET status = 'pending', worker_id = NULL
WHERE status = 'processing'
  AND claimed_at < NOW() - INTERVAL '5 minutes';

No advisory locks. No Redis. Pure PG. Works.

You

/caveman:compress CLAUDE.md

My CLAUDE.md is getting long and I want to save tokens on every session start.

Agent

Compressing CLAUDE.md → CLAUDE.original.md (backup) + CLAUDE.md (compressed).

Reading file...

Original: 847 tokens

Compressing — rules intact, prose trimmed, code/paths/commands untouched.

Compressed: 341 tokens Saved: 60%

CLAUDE.md          ← Claude reads this (341 tokens each session start)
CLAUDE.original.md ← you read and edit this (847 tokens, human-readable)

What changed:

"Make sure to always use pnpm and never npm" → "Use pnpm. Never npm."
"When you are writing tests, you should make sure that..." → "Tests: use supertest. Mock external only with msw. Always test 400/500."
"It is important that commits do not include Co-Authored-By lines" → "No Co-Authored-By in commits."
All code blocks, file paths, URLs, commands: untouched

Workflow going forward:

Edit CLAUDE.original.md (readable)
Run /caveman:compress CLAUDE.md to regenerate compressed version
Claude reads the compressed version — 60% fewer input tokens every session

Run again after any CLAUDE.md update. Compress is idempotent — safe to re-run.

Get Started

How to install and use this skill in your preferred environment.

Skills are designed for AI coding agents (Claude Code, Cursor, Windsurf) and IDE-based workflows where the agent can read files, run scripts, and act on your codebase.

Models & Context

Which AI models and context windows work best with this skill.

Recommended Models

Works on all models — the skill reduces output tokens regardless of model capability. Biggest absolute savings come from verbose models (Opus, GPT-4o) where the uncompressed output is longest. Smaller models (Haiku, GPT-4o-mini) produce shorter output already so savings are proportionally lower.

Context Window

The SKILL.md is tiny (~2KB). Negligible context overhead — this is the point.

Pro tips for best results

1

Be specific

Include numbers — users, budget, RPS — so the skill can size the architecture.

2

Share constraints

Compliance needs, team size, and existing stack all improve the output.

3

Iterate

Start with a high-level design, then ask follow-ups for IaC, cost analysis, or security review.

4

Combine skills

Pair with companion skills below for end-to-end coverage.

Works Great With

These skills complement Caveman for end-to-end coverage. Install them together for better results.

Self-Improving Agent

Curate auto-memory into durable project knowledge, promote patterns to rules.

AgentMemory

Git Commit Hygiene

Install a commit-msg hook that auto-strips AI tool attribution from git commits.

GitHooks

PR Review Expert

Blast radius analysis, security scan, and coverage delta for PRs.

PR ReviewSecurity

Code Reviewer

Code review automation for TypeScript, JavaScript, Python, Go, Swift, Kotlin.

ReviewQuality

$ skillfish add johnefemer/skillfish --all # install all skills at once

Good to Know

Advanced guide and reference material for Caveman. Background, edge cases, and patterns worth understanding.

Contents

Output tokens vs. input tokens

Caveman cuts output tokens — what the model writes back to you. It does not affect input tokens (what you send or what gets loaded from CLAUDE.md, memory files, etc.).

For input token savings, use caveman-compress: it rewrites your CLAUDE.md and memory files into caveman-speak so Claude reads less at session start (~46% savings per file). Both together give you the full picture.

Technical accuracy is not compressed

The skill explicitly preserves: technical terms, error messages (quoted exactly), code blocks, file paths, URLs, version numbers, and warnings. Only prose fluff is removed. A March 2026 paper found that brevity constraints on LLMs can improve accuracy by 26 percentage points on some benchmarks — less rambling, more precise answer.

Auto-clarity for dangerous operations

Caveman automatically suspends compression for:

Destructive operations (DROP TABLE, rm -rf, irreversible actions)
Security warnings
Multi-step sequences where fragment ordering risks misread

After the clear part is done, caveman mode resumes automatically. You don't need to toggle it.

Intensity levels

Level	Character count vs. normal	Best for
`lite`	~30% savings	Documentation, specs, design reviews
`full`	~65% savings (default)	Daily coding, debugging, Q&A
`ultra`	~80% savings	Quick lookups, log triage, you know the domain
`wenyan`	~80–85% savings	Curiosity, show off, token art

Level persists until changed or session ends. Set once, forget.

Works across all agents

The skill was designed for Claude Code but the same SKILL.md content works in Cursor, Windsurf, Copilot, Cline, Codex, Gemini CLI, and 40+ agents via npx skills add. For agents without a hook system (most of them), caveman won't auto-start — say /caveman or "talk like caveman" each session, or paste the always-on snippet into your agent's rules file.

Claude Code: auto-activate every session

Install once and caveman activates automatically at session start via the SessionStart hook in the plugin. No need to type /caveman each time.

The benchmark numbers are real

The repo includes a reproducible eval harness (evals/) with a proper three-arm control — not just "verbose vs. caveman" but also "terse vs. caveman" — to isolate the skill's contribution from generic terseness. You can run it yourself with the Claude CLI.

Ready to try Caveman?

Install the skill and start getting expert-level guidance in your workflow — any agent, any IDE.

$npx skills add JuliusBrussee/caveman

← Browse all skills

Caveman

What This Skill Can Do

Real Conversations

Get Started

Claude Code, OpenCode, or any CLI agent

Cursor, Windsurf, or IDE-based agents

Claude.ai, ChatGPT, Gemini, or any web AI

Models & Context

Recommended Models

Context Window

Pro tips for best results

Works Great With

Self-Improving Agent

Git Commit Hygiene

PR Review Expert

Code Reviewer

Good to Know

Output tokens vs. input tokens

Technical accuracy is not compressed

Auto-clarity for dangerous operations

Intensity levels

Works across all agents

Claude Code: auto-activate every session

The benchmark numbers are real

Ready to try Caveman?