caveman

# Caveman ## Overview Strip filler, articles, pleasantries. Keep full technical accuracy. Target ~75% token reduction. ## When to Use - Writing sub-agent prompts or persona handoff messages - Producing rubber-duck reports for internal pipeline consumption - Any agent-to-agent communication where token reduction matters ## Scope restriction Doc-only workflow. Agent-to-agent communication ONLY — sub-agent prompts, rubber-duck reports, persona handoffs, internal pipeline messages. ## Why not human-facing? - **`enforce-npov`** — due weight requires full attribution register with proper natural language; telegraphic style drops necessary hedging. - **`detect-ai-tells`** — uniform telegraphic cadence is a known AI-generation artifact; triggers false positives. ## Rules 1. Drop articles (a/an/the). 2. Drop filler phrases ("it should be noted", "in order to", "it is important"). 3. Drop pleasantries ("please", "thank you", "I'd be happy to"). 4. **Keep** all technical terms exactly. 5. **Keep** all proper nouns, identifiers, paths, and numbers. 6. **Keep** all SourceRef citations verbatim. 7. Preserve meaning exactly — compression must not alter semantics. ## Verification Compare original output vs compressed output: 1. Count tokens (approximate: split on whitespace). Target: compressed ≤ 25% of original token count. 2. Verify all technical terms, proper nouns, paths, numbers, and SourceRef citations are present in the compressed output. 3. If token reduction is < 50%, re-apply rules — the output likely retains too much filler. > Adapted from mattpocock/skills `caveman`. KB variant restricts to agent-to-agent per NPOV/AI-tells policy.