caveman hit 97,000 stars for a token-cutting skill, and its own README publishes the test that drops the claim from 65% to 8.5%.

caveman is MIT-licensed, makes coding agents drop prose filler while leaving code, commands, and errors byte-for-byte identical, installs in one command across 30+ agents with six compression levels. The headline 65% comes from 10 chat-style prompts with a 22–87% range. The README also carries an independent JetBrains test across 86 real coding tasks measuring 8.5%, because most agentic output is code, not narration. It further discloses that the skill only shrinks output tokens, adds ~1–1.5k input tokens per turn, and can be net-negative on terse workloads. That's the most honest accounting I've seen in a viral repo. Plan against 8.5%.