Caveman Prompt Cuts Claude Tokens 45% via Filler Stripping
Caveman skill drops articles, filler, hedging from Claude outputs for 45% fewer tokens vs baseline (39% vs 'be concise'), netting 39% cost savings on follow-ups despite higher input costs.
Enforce Concise Outputs by Dropping Filler and Hedging
Caveman applies strict rules to Claude prompts: drop articles (a/an/the), filler words (sort of, basically), pleasantries (thanks, please), and hedging (might, possibly). Use short synonyms (big for extensive, fix for implement). Preserve technical terms, code blocks, errors. Structure responses as thing → action → reason → next step. This transforms verbose explanations—like a Next.js auth demo from multi-sentence prose to bullet-point flows (e.g., "app load → check localStorage → fake user")—delivering technical info without readable English fluff.
Test on 10 prompts (e.g., "Git rebase vs merge") shows 45% output token reduction vs baseline Claude, 39% vs just prompting "be concise." Baseline: ~8¢ output; Caveman: ~4¢. Input jumps to 4¢ due to skill's Markdown file, making single prompts 10% pricier overall. But follow-ups hit prompt cache, flipping to 39% net savings since cached input costs less.
Boost Accuracy 26% with Brevity Constraints
Constraining LLMs to brief responses improves technical accuracy: a 2024 study found 26% gains on benchmarks. Caveman's terse style mimics this, prioritizing signal over politeness—e.g., arrows for flow (load → check → login) cut reading time while retaining precision. Install via Vercel AI SDK: npx @vercel/ai-sdk@latest add skills.sh/juliusbrussee/caveman. Default "full" mode balances brevity; tune with light/ultra intensity (ultra abbreviates, strips conjunctions, one-words-only).
Specialized Modes for Commits, Reviews, Compression
Wenyan mode uses token-efficient classical Chinese (unreadable for most). Caveman Commit: terse conventional commits (e.g., "fix: auth flow"). Caveman Review: one-line code findings. Compressed: Caveman-ify input files to shrink natural language before reuse, trimming input tokens further. Use for code analysis, docs, or any verbose LLM task where facts > prose.