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
Search the web, scrape URLs into markdown/HTML/screenshots/links/images, extract structured JSON with AI prompts, crawl multi-page sites, and monitor pages for changes on a schedule. Supports browser automation features like JavaScript rendering, stealth mode, cookies, and custom headers. Perfect for data engineers, researchers, and developers who need API-free web data collection without writing scrapers.
This comprehensive collection includes 79 dedicated agent skills for all major Azure services and operations. You can automate Azure resource deployment and management, query infrastructure and logs with natural language using Kusto Graph and IRQL, analyze security events and detect threats, configure identity and access management through Entra ID, monitor application performance with Application Insights, manage storage and messaging services, execute cloud migration workflows, validate compliance policies, and query Azure resources by metadata or properties.
This comprehensive skill collection guides you through creating professional videos programmatically using Remotion and React. Create new video projects, structure complex multi-scene compositions, and animate content with full interactivity. Add captions by transcribing audio, importing SRT files, or generating them programmatically. Build animated maps with Mapbox, MapLibre, or GeoJSON, including 3D geographic flyovers. Process multimedia by trimming, cropping, and extracting metadata from video and audio files.
Render dynamic videos directly from HTML/code using Puppeteer and ffmpeg, with built-in support for AI voiceover, caption overlays, and motion animation. Convert markdown changelogs into polished 45-60 second branded videos with animated UI mockups, captions, and AI narration. Create explainer videos, product demos, and animated visualizations using code-first workflows and motion doctrines that enforce consistency and quality.
by heygen-comRigorPilot is a comprehensive skill suite for deep learning research, providing auditable repository analysis, README-first reproducibility workflows, bounded exploration with evidence-based ranking, and publication-ready scientific documentation. Each skill enforces rigor through documented assumptions, comparability tracking, and explicit decision checkpoints. Use these skills to understand ML repository internals, faithfully reproduce published results, systematically explore novel ideas with proper baselines, and generate standardized research artifacts.
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.
Present users with an adaptive 24-question personality assessment that uses Bayesian active learning to efficiently map their psychological profile. The skill automatically maps user responses (1-7 scales or natural language) to archetype classifications based on a 5-color psychological model, tracks probability distributions in real-time, and delivers detailed personality insights including core strengths, weaknesses, and top archetype matches.
This 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.