I measured the viral "talk like a caveman" token saver. A two-word prompt got most of it.
Description and use
Measured token savings with Caveman: 63.6% in ultracave mode, 42.9% in caveman mode, 28.7% with "Be concise.". All outputs compile and pass the test, maintaining quality.
Problem
Assess token savings and quality impact of the Caveman repo’s skill and proxy for LLMs.
Outcome
The /ultracave mode achieves the advertised 65% token reduction, while the /caveman skill saves ~43% and "Be concise." saves ~29%. All returned code compiles and passes the original test, confirming quality is maintained.
Workflow
- 1. Identify a real C# bug with a verifiable pass/fail.
- 2. Prepare four prompt variants: plain, "Be concise.", /caveman, /ultracave.
- 3. Run each variant three times, counting tokens with o200k_base tokenizer.
- 4. Compile returned code against original Cart and run test.
- 5. Compare token counts and quality results.
Evidence
The description says 65%, from talking like a caveman.
Install the skill for shorter answers, or the CLI proxy to cut input tokens by 33.2%.
All four compiled. All four printed PASS.
- Record date: 2026-10-16T11:41:34.405Z