JuliusBrussee/caveman
🪨 why use many token when few token do trick. Viral skill + proxy for coding agents that cuts 65% of tokens by talking like a caveman.
Read original source
Source record
Recorded metrics
daily_stars · 227 stars today
2026-11-15 04:11:08.796Z
GitHub daily Trending window; attention only
Project analysis
Problem & scenario
The supplied evidence describes Caveman, not a verified WeWorkBuddy product: it targets coding-agent users whose sessions consume tokens through verbose replies and bulky tool output. Example scenario: a developer investigates a failing test. Before, the agent reads extensive test logs and produces a long explanation; with Caveman, supported output is compressed before the model reads it, and reply rules shorten the explanation while preserving code and errors verbatim. This makes token reduction relevant to metered usage, but shorter replies alone save nothing under per-request billing. Whether WeWorkBuddy incorporates this product is unknown.
Features & advantages
Caveman extends beyond a brevity instruction: its documented scope includes compressing tool results, selecting model effort through Auto, and reusing scripts through Blocks. Against the named baseline “Answer concisely,” the README reports an additional 3% median output-token reduction for /caveman and 35% for /ultracave in an older evaluation. These are author-reported results, not a demonstrated general advantage over Codex or Claude Code. The practical choice is between a lightweight reply skill and a broader proxy setup. Compression is workload-dependent: the reported HTML session used more tokens, and the voice rules themselves add input tokens.
Technical approach
The README documents a local proxy workflow that reduces supported tool output before it reaches the model. A representative documented flow is: an agent obtains logs, JSON, CSV or test output; Caveman compresses that material; the model receives the reduced representation; the original remains on disk for retrieval. Keeping originals provides a recovery route when details are needed, but compression coverage limits the benefit: the reported suite had no HTML compressor. Installation requires Node.js 22.13+ and uses existing provider credentials. Auto adds account-dependent model selection. No commit-pinned source code was supplied, so exact interception, compression and retrieval functions cannot be traced or verified.
Business model & buyers
The documented offer is free software used with subscriptions and API keys users already have; Auto also requires a free account. Free access to Caveman does not establish free underlying model usage. The likely adopter is a developer already paying for coding agents, with a purchase or installation trigger of high metered usage or repeated bulky tool output. That buyer profile is an inference, not customer evidence. Per-request billing weakens the financial case for reply shortening, while broader compression or routing needs workload-specific evaluation. The supplied README gives no paid Caveman tier, checkout terms or revenue evidence; hypothetical monetization should not be treated as an existing business model.
Growth & distribution
GitHub discovery and a one-command installation provide a plausible adoption path, but the evidence does not establish sustained growth. S1 records a GitHub Trending listing; S2 claims additional exposure through Hacker News, research mentions and a creator video. These are distribution signals, with no dated traffic or conversion series supplied. A developer can first try the reply-only skill, experience shorter answers, then adopt the local proxy and optional Auto account. That progression is inferred. Activation depends on compatible agents and, for the CLI, Node.js 22.13+; retention likely depends on useful savings in the user's own workload. Rankings and reported benchmarks do not establish repeat usage.