Exploring a fascinating new project that cuts AI output by 75% using caveman-speak, supports 40+ coding agents, and even includes classical Chinese mode.
And hey, quick heads up - after this episode, check out our bonus episode where we actually explain Caveman using caveman-speak itself. It is hilarious and meta. But first, let me tell you about a project that's only 10 days old but already supports over 40 different AI coding agents.
Welcome to Code2Cast! I'm Chris, and today I'm excited to talk about a project that's only 10 days old but already supports over 40 different AI coding agents. Jessica, meet 'Caveman' by Julius Brussee.
Oh this is wild! So instead of Claude saying 'I'd be happy to help you with that. The issue you're experiencing is likely caused by your authentication middleware not properly validating the token expiry' - it just says 'Bug in auth middleware. Token expiry check use less-than not less-than-equals. Fix:' That's it?
Exactly! And the crazy part is it's not just a gimmick - Julius has benchmarked this with real API calls. We're talking 65-75% token reduction on average, with some responses hitting 87% savings. Same technical accuracy, just no fluff.
But what really blew my mind is the architecture. This isn't just a simple prompt - it's a full ecosystem. For Claude Code, they built this sophisticated hook system with SessionStart hooks that auto-inject the caveman rules, UserPromptSubmit hooks tracking mode changes, even a statusline badge showing your current caveman level.
The distribution strategy is insane. Look at this - they have a GitHub Actions workflow that takes one source file and auto-generates agent-specific versions for Cursor, Windsurf, Cline, Copilot, Gemini CLI, Codex... Each one has different frontmatter requirements, different hook systems, different installation methods.
And the intensity levels! You can go 'lite' for professional terseness, 'full' for classic caveman, or 'ultra' for maximum compression. But here's the kicker - they also have 'wenyan' mode. That's classical Chinese! Julius discovered that classical Chinese is literally the most token-efficient written language humans ever invented.
The engineering behind this is remarkable. They have this 'auto-clarity' system that detects when caveman mode might be dangerous - like for security warnings or destructive operations - and automatically switches back to normal prose, then resumes caveman mode after.
What gets me is how rigorous the validation is. They built a three-arm evaluation harness - not just 'verbose vs caveman' but 'baseline vs terse vs caveman' to prove the compression isn't just from being less verbose. They're running real benchmarks with the Claude API, tracking actual token counts, not estimates.
And they even built a compression tool that works on input tokens! Your CLAUDE.md file that loads every session? It compresses that to caveman-speak too, saving 46% on average. So you get both input and output token savings.
The timing is perfect too. With all these AI coding agents becoming mainstream - Cursor, Claude Code, Windsurf, Copilot Chat - everyone's dealing with verbose AI responses. Julius basically said 'why use many token when few do trick' and built a production system around it.
For a 10-day-old project, this shows incredible systems thinking. Julius didn't just build a skill - he built an entire distribution platform that adapts to the quirks of 40+ different agents. That's the kind of engineering that makes me think this could become the standard way people interact with AI tools.
Absolutely. When you see responses like 'Inline obj prop arrow new ref arrow re-render. useMemo.' instead of three paragraphs explaining React re-renders... you realize how much cognitive overhead we've been accepting from verbose AI. Caveman just cut straight to the answer.
That's Caveman by Julius Brussee - a fascinating example of how the right constraint can completely change how we think about AI interaction. Links in the show notes, and we'll definitely be following this one as it evolves.