A tool that makes AI respond shorter, helping you read faster and save time — without losing any important technical details.
Ever had a long AI session where tokens pile up faster than you can keep track? Or found yourself reading through walls of filler words in commit messages and review comments? Caveman is a set of 6 skills designed to solve exactly that — reducing the words you need to read while keeping all important technical data intact.
Core concept: Caveman is not about being rude — it's about information compression. Think of it like texting a close friend: short, to the point, no need for elaborate greetings.
Caveman is a collection of skills that reduce unnecessary word count in AI-assisted workflows, while not losing critical technical data. The result? You save input tokens, reduce reading time, and maintain high-precision reading of code, commands, and error strings.
⚠ Without Caveman
Hi, I think there seems to be a small error around line 42 in this code, maybe the variable x wasn't initialized properly, could you check it out please?
✅ With Caveman
file:L42: var x not initialized. Init = null.
The suite includes 6 components, each serving a specific purpose:
| Skill | Type | Function |
|---|---|---|
caveman |
Persistent mode | Compress output per benchmark while preserving technical accuracy |
caveman-commit |
One-shot | Generate short, precise Conventional Commit messages, prioritizing "why" |
caveman-compress |
File mutation | Compress prose in memory files to reduce input tokens |
caveman-help |
Reference | Show quick-reference for modes, skills, commands, and configuration |
caveman-review |
Review output | One-line review comments: location, problem, fix |
caveman-stats |
Hook-backed | Show token usage, savings, rule overhead, and session net |
/caveman # Full mode (default) /caveman lite # Lite mode /caveman ultra # Ultra mode /caveman wenyan-lite # Wenyan (classical Chinese) light /caveman wenyan-full # Wenyan full /caveman wenyan-ultra # Wenyan maximum /caveman off # Disable, return to normal mode
⚠ Lite
Remove filler, keep full sentences.Error: var x not initialized in file L42.
✅ Full
Default; fragments allowed.file:L42: x not initialized.
⚠ Ultra
Maximum compression, no self-made abbreviations.
✅ Ultra
L42: x undef.
This command generates short, precise Conventional Commit messages:
# Run command: /caveman-commit or /commit # Output: feat(auth): fix null ref in login flow
Rules: Keep subject to max 50 characters when possible. Only add body when necessary. Hard cap at 72 characters.
Generate one-line review comments in this format:
file:L42: severity: problem. Fix.
Example:
auth.py:L87: high: missing null check. Add guard.
Compress prose in memory files to reduce input tokens:
/caveman-compress /path/to/memory.md
Safety mechanism: Creates backup before overwriting, validates after compression, and retries up to 2 times. Not for code, JSON, YAML, HTML, CSS, SQL, or lock files.
ALWAYS PRESERVE:
PAUSE compression when:
| Criteria | Without Caveman | With Caveman |
|---|---|---|
| Response length | 150-300 words | 10-50 words |
| Commit message | "Fix bug in auth" | feat(auth): fix null ref in login |
| Review comment | "Could you please check if..." | auth.py:L87: missing null check. |
| Memory file | Long prose, hard to scan | Auto-compressed, safe backups |
| Token usage | Piles up fast | Significantly reduced, measurable |
If you're new to Caveman, start with /caveman lite, then move up to /caveman (full). Ultra mode is only when you're comfortable.
Caveman automatically knows when to "slow down" — keeping full sentences when content relates to security or critical actions. You don't need to micromanage every detail.
Use /caveman-stats on compatible hosts to see real token savings data. Don't guess — let the hook compute from session logs.
Caveman is not a gimmick — it's a tool to work more efficiently with AI. When you reduce the words you need to read, you free up mental energy for what truly matters: reading code, solving problems, and making the right decisions.
💡 In summary: