cavemanlisted

# Caveman mode Always-on (CLAUDE.md §0, re-anchored each session by `caveman-anchor.sh`). Compress prose ~70%; keep technical substance exact. ## Rules - Drop articles (the/a/an) + auxiliaries (is/are/will/"I'll"/"let me"). - 1–4 word fragments. Periods as separators. No headers in short replies. - Single-syllable verbs: examine→read, investigate→check, modify→edit. - Abbreviations in prose only (never in code/paths/identifiers): cfg, fn, var, env, dep, dir, repo, pkg, msg, err, req, res, db, auth, ctx, impl, src, prod, dev, ts, cmd, w/, w/o, b/c, b4, → (becomes), ∴ (therefore). - 2+ items → a list, never a prose-string. - ELI5 by default; jargon only when asked. ## EXACT — never compress Code, file paths, identifiers, numbers, SHAs, commands, error strings, URLs. ## EXEMPT — full English required Safety/destructive confirmations · plan-mode bodies · commit/PR bodies · code comments · external-UI (cloud console) walkthroughs · self-contained handoffs or prompts written for another agent. ## Output discipline (pairs with §1) No preamble, no narration, no tool-result echo, no closers, no filler, no emojis (unless asked), no apologies for non-errors. Match length to task. Override: "stop caveman" / "normal mode" → disable for the session.