Caveman Mode
Cut token usage 65% by talking like a caveman
What You Can Do
Drastically reduce token consumption in Claude interactions by using ultra-compressed communication that preserves full technical accuracy. Caveman mode strips articles, filler, and hedging while keeping all technical substance, code, and meaning intact. Choose from multiple intensity levels (lite, full, ultra) and linguistic variants (wenyan-lite, wenyan-full, wenyan-ultra). Includes cavecrew subagent delegation that compresses agent output by 60%, memory file compression for CLAUDE.md files, and real-time token savings tracking.
Features
Strip articles, filler, and hedging while preserving technical accuracy. Measured 65% token reduction with full substance retention.
Choose from lite (professional tight), full (classic caveman), ultra (stripped conjunctions), or wenyan variants for classical Chinese compression.
Spawn caveman-compressed investigator, builder, or reviewer agents. Tool results return 60% smaller, extending main context duration across long sessions.
Compress CLAUDE.md, todos, and preferences while preserving code blocks, URLs, file paths, and all technical content exactly.
Real-time stats showing actual token savings, estimated overhead, and net efficiency gains with honest warnings for negative-savings scenarios.
Preserves user's dominant language and technical terms verbatim. Compresses style, not language, with classical Chinese mode.
Active every response until explicitly turned off. No filler drift. Auto-triggers when token efficiency requested.
Example Output
Extending session context
How can I run longer Claude sessions without losing context?
Response: Three ways. Lite mode drops filler but keeps grammar. Cavecrew split work to subagents with 60% smaller results. Compress CLAUDE.md with /caveman-compress saves 40-60% of memory file.
High-volume token optimization
We run 100 Claude analyses daily and token costs are high.
Response: Caveman full mode in batch scripts. Measured 65% output token savings, zero accuracy loss. Use cavecrew subagents for investigation phases. Compress memory at session start. Backend savings: ~$30-40 per 1000 analyses.
Mode intensity comparison
Show me the same technical answer in different caveman intensities.
Response: Full mode: "Component re-renders because new object ref created each render. Wrap in useMemo." Ultra mode: "New object ref each render triggers re-render. useMemo fixes." Wenyan-full: "每繪新生對象參照，故重繪；以 useMemo 包之則免。"
What's Included
- Caveman skill: Main mode with full, lite, ultra, and wenyan intensity levels. Auto-activates on user request or token-efficiency context.
- Cavecrew system: Three subagent presets: cavecrew-investigator (locate code), cavecrew-builder (1-2 file edits), cavecrew-reviewer (diff audit). All output compressed.
- Caveman-compress utility: Compress natural language memory files while preserving code blocks, URLs, paths, and technical content exactly. Creates human-readable backups.
- Caveman-stats tracker: Real-time token usage metrics, estimated savings, per-turn overhead, and net efficiency with honest warnings for negative-savings scenarios.
- Decision guides and documentation: When to use caveman vs vanilla agents, chaining patterns (locate-fix-verify), and compression rules that guarantee accuracy preservation.
Who It's For
- Claude Code users managing context limits
- Engineering teams optimizing token budgets
- Researchers running high-volume LLM analysis
- Developers in long multi-hour coding sessions
Best For
- Extending main context duration in long sessions
- Reducing token costs for batch processing
- Maintaining accuracy while cutting verbosity
- Memory file compression and archival
You might also like
Create production-ready mobile app designs by describing what you want in plain language. Use chat-based requests to design screens, iterate with follow-up messages, and export code implementations for HTML, React Native, and SwiftUI. Features include live design preview in the Sleek editor, component screenshot generation, and built-in design references to seed your visual style.
Access 29 production-ready Claude agent skills that wrap RunComfy's AI models for image generation, video creation, audio synthesis, and avatar videos. Skills include music generation (ACE Step with inpaint/outpaint), AI avatars and lip-sync (OmniHuman, Seedance, HappyHorse), image editing (inpainting, outpainting, face-swap, relight), video generation and extension, and upscaling. All skills are CLI-driven via the runcomfy command, with built-in schema validation, error handling, and examples for each model.
Elevate your interface design from strategy to pixel-perfect execution. Shape layouts and user flows before coding, critique designs against established heuristics, audit for accessibility and performance gaps, and polish for production. Enhance with purposeful animation, strategic color, harmonious spacing, and delightful moments. Includes comprehensive guidance for websites, dashboards, apps, components, forms, onboarding, and design systems across web and native platforms.
by pbakausThis collection of skills improves code quality, architecture documentation, and team workflow coordination. Review Architecture Decision Records with fresh perspective and rewrite them for clarity, implement type-safe AST visitor patterns for robust code organization, automate Biome linting upgrades across monorepos, and generate well-structured GitHub pull requests integrated with Linear project tracking. Perfect for teams that care about maintainable code, clear decision documentation, and smooth development workflows.
Generate batches of proven viral hooks across named-creator frameworks (MrBeast, Hormozi, Brunson, Koe, Welsh, Bush, Galloway). Write and audit short-form scripts with platform-specific pacing. Optimize Instagram Reels for sends-per-reach using Trial Reels testing. Craft captions and CTAs that match video payoff. Diagnose why openings flop and rewrite weak hooks against anti-patterns. The skills encode what patterns tend to lift retention in the first 1 to 3 seconds, then repurpose across platforms.
Tackle specialized engineering problems with expert Claude skills spanning CI/CD automation, infrastructure security, open source governance, browser scripting, and documentation workflows. Get authoritative answers to GitHub Actions, Linux hardening, and project sustainability questions grounded in official docs and best practices. Automate routine tasks like userscript development, README localization, and configuration management across diverse technical domains.
Discover Reddit threads where people are genuinely shopping for what you sell, then draft replies that sound like a helpful community member (not an ad). The skill runs two moves: discovery scans target subreddits for recommendation asks, expressed pain, and competitor mentions, ranking threads by signal, product fit, and timing; drafting writes experience-based replies that respect community norms, disclose your affiliation honestly, and only mention your product when it truly answers the question.
Design and build animations using a proven decision sequence covering when to animate, which tools to use (CSS, WAAPI, Motion.dev), property selection, easing curves, and accessibility. Review existing animations and audit codebases for quality issues. Understand and name motion effects using a comprehensive vocabulary glossary. Learn from Emil Kowalski's opinionated motion design philosophy built on decades of interface engineering experience.