Labs / research
The caveman phenomenon
Abstract
Caveman is a skill file that instructs a coding agent to answer like a caveman: drop the articles, drop the filler, keep every technical fact, and never touch code, commands, or error strings. The joke installs from one file into more than thirty coding agents; it has passed 72,800 GitHub stars and reached number one on Hacker News. Underneath the joke sits a measurable claim about the price asymmetry of LLM output. Across a ten-prompt benchmark the skill reduced output tokens by 65 percent on average, with a range of 22 to 87 percent depending on task mix. We describe the mechanism, report the measurements with their limits, and present CaveGemma, a fine-tune of google/gemma-4-31B-it that reproduces a weaker form of the register with no prompt in context at all: 27 percent fewer output tokens across 193 held-out pairs, with code-fence exactness between 96.3 and 100 percent.
1. Background
Output tokens are the expensive side of the meter. Providers price them above input tokens, and none of the caching machinery that discounts repeated input applies to them; every word the model says is billed at full rate, every time. Most of what an assistant says is also style. The hedging, the restating of the question, the "certainly, happy to help" framing: none of it is the answer.
Together those two facts make assistant style a pure cost. Caveman deletes it. People shared the skill because it is funny; people kept it because the bill went down.
2. The intervention
The skill is a single markdown file the agent reads at session start. It instructs terse, fragment-based answers and enumerates what may never be compressed: code blocks pass through unchanged, commands stay exact, error messages are quoted verbatim. Compression lands on prose and only on prose.
The same file installs into Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, and more than thirty other agents. It ships four intensity levels: lite, full, ultra, and wenyan, the last of which answers in classical Chinese, the densest register we found that current models still write fluently.
3. Measurements
| Quantity | Value |
|---|---|
| Benchmark size | 10 prompts |
| Mean output-token reduction | 65% |
| Range across prompts | 22–87% |
| Scope | Output tokens only |
| Code, commands, errors | Excluded from compression by instruction |
The mean is 65 percent, but the honest figure is the range. A prose-heavy explanation compresses toward the top of it; a code-heavy answer compresses toward the bottom, because code passes through untouched and the skill has nothing left to delete. Savings depend on task mix, and anyone quoting the average alone is selling something.
Two scope boundaries matter. The benchmark covers output tokens only; input and reasoning tokens are unchanged, and the skill itself adds prompt overhead on the input side. Second, this is compression of style, never of meaning. Where technical context does get structurally compressed elsewhere in the stack, in the engine's JSON, log, AST, and tool-schema compressors, the original bytes are kept in the CCR store and recovery is byte-exact.
4. From prompt to weights: CaveGemma
A prompt that saves tokens is a trick; a model that cannot help saving them is a property. CaveGemma is a QLoRA rank-16 fine-tune (Dettmers et al., 2023) of google/gemma-4-31B-it trained to write the caveman register natively, with no skill file in context. It ships as merged weights (62.5 GB) or a 534 MB LoRA adapter and loads through Hugging Face transformers.
| Quantity | Value |
|---|---|
| Held-out evaluation pairs | 193 |
| Output-token reduction (weighted mean) | 27% |
| Code-fence exactness | 96.3–100% |
| Semantic similarity to base answers | 0.91–0.98 cosine |
| Article density (share of a, an, the) | 0.005–0.020, vs. 0.080 in sources |
| Evaluation modality | Text only |
The two headline numbers are not comparable. The skill's 65 percent comes from a ten-prompt suite with the instruction sitting in context; CaveGemma's 27 percent comes from 193 held-out pairs against the base model with no instruction at all. The fine-tune buys less compression than the prompt does, and it buys it everywhere, unconditionally, without spending input tokens to ask.
5. Limitations
Both suites are small: ten prompts for the skill, 193 pairs for the fine-tune. The skill's code-exactness guarantee is an instruction the agent follows, not a measured property; the fine-tune's exactness is measured and imperfect, since 96.3 percent on the worst pair set means it can still damage a code fence. Article density is a proxy for register adoption, not for answer quality; the semantic-cosine band of 0.91 to 0.98 is the closest we have to a quality retention measure, and cosine similarity is a blunt one. CaveGemma's multimodal behavior remains untested.
Every number in this report is an inferred measurement on a benchmark suite. None of it is a verified-savings claim in the ledger sense (see The zero-dollar dashboard).
References
- Dettmers, T., Pagnoni, A., Holtzman, A., and Zettlemoyer, L. (2023). QLoRA: Efficient Finetuning of Quantized LLMs. NeurIPS 2023.
- google/gemma-4-31B-it. Model card, Hugging Face.