An ultra-terse overlay that keeps the substance
"Sure! I'd be happy to help you with that." Real technical substance shows up buried under pleasantries, hedging, and articles on every single response. Caveman is a loadable Amplifier behavior overlay that strips it.
And that padding isn't free — it has a price on every turn.
The bundle self-describes as an "Ultra-terse LLM communication overlay. ~75% fewer output tokens." It's a claimed design target — no benchmark file in the repo measures the exact figure.
Cutting that much sounds great — until you ask what gets cut.
Anything can cut words. The risk is cutting the technical payload you actually needed — fragments can mangle meaning, and fragment order can risk a misread.
So the rule has to be surgical: remove fluff, never substance.
One precise rule from caveman-common.md: strip articles, filler, pleasantries, and hedging — while technical terms stay exact, code blocks stay unchanged, and errors are quoted exact.
One rule, three intensities — here's the dial.
Caveman ships lite, full (the default), and ultra. The whole bundle is 8 tracked source files totaling 105 lines — small enough to read and trust in a sitting.
Small and terse — but what stops it from being terse at the worst moment?
Caveman's Auto-Clarity carve-out drops terseness for security warnings and irreversible-action confirmations — those render in full normal prose. Terseness resumes after the clear part is done.
Judgment about when to stop is what makes it trustworthy, not just clever.
Drop caveman for: security warnings, irreversible action confirmations, multi-step sequences where fragment order risks misread, or when the user asks to clarify. Resume caveman after clear part done.
Caveman is active every response — no revert after many turns, no filler drift, still active if unsure. It's off only when you say "stop caveman" / "normal mode" or disable via /config.
Opt into terseness once; the safety stays automatic.
Strip the fluff. Keep the substance. Let the tool know when clarity beats brevity.
Caveman v1.0.0 is a tiny, loadable Amplifier overlay layered on microsoft/amplifier-foundation — ~75% fewer output tokens by design, with the technical payload and the safety carve-out kept intact.
Released — bundle v1.0.0
Data as of: 2026-04-15 (last commit). Subject: bkrabach/amplifier-bundle-caveman, identified by cloning and reading the primary-source repo directly.
Commands run:
git remote -v → git@github.com:bkrabach/amplifier-bundle-caveman.gitgrep -rn '75' . --include='*.md' → only bundle.md description; no benchmark filefind . -not -path './.git/*' -type f | xargs wc -l | tail -1 → 105 total; git show 64e85db --stat → 8 files, 79 insertionsls behaviors context; cat bundle.md; grep -A3 include behaviors/caveman.yamlcat context/caveman-common.md; cat context/caveman-full.mdgit rev-list --count HEAD → 2; git shortlog -sne HEAD → 2 commits, author "Amplifier Bundle"Gaps: The ~75% figure is the bundle's own description string, not an independently measured result — treat as a claimed design target. "100% technical substance intact" is design intent in the context files, not a measured guarantee.
Contributors: Sole commit author "Amplifier Bundle <caveman-bundle@amplifier.local>" (2 of 2 commits); repo owner bkrabach per remote.