Caveman Transfer

Purpose

Caveman makes the model talk less. caveman-transfer makes the prompt smaller before sending. Different sides of the bill. They compose.

The skill calls transfer.tokenstree.com — a hosted LLMLingua-2 (xlm-roberta-large fine-tuned for prompt compression) — and returns the semantically compressed text. LLMLingua-2 drops low-information tokens while preserving meaning. Independent benchmark: -52.9% average input tokens across the caveman benchmark suite, with code/URLs/proper nouns preserved.

Trigger

/caveman-transfer <filepath> or /caveman-transfer "<text>" — or invoke when:

• User pastes a long prompt and asks "make this smaller"
• Memory/CLAUDE.md/RAG context exceeds ~2k tokens
• User says "compress prompt", "compress context", "shrink input"

Process