JuliusBrussee/caveman

A cross-agent skill that strips filler from Claude’s replies and answers in terse, token-lean prose while keeping code, commands, and error text byte-for-byte exact.

90.3k stars 5.1k forksLicense: MITUpdated 2026-07-03
Productivitytokensoutputefficiencycross-agent
View on GitHub

Why it matters

Verbose model output is a hidden tax — every restated question and hedging paragraph costs tokens and attention. Caveman clamps responses down to the essential answer without touching literal code or commands, so you pay for signal instead of padding. On a heavy day of agent use the savings are continuous, which is why it spread far beyond Claude to thirty-plus other tools.

Use cases

  • Cut output-token spend on high-volume agent sessions
  • Keep long conversations readable by removing hedging and filler
  • Preserve exact code and shell commands while trimming prose
  • Apply one terse-output rule consistently across many agents

Quickstart

npx skills add JuliusBrussee/caveman

Related repositories

OthmanAdiProductivity

planning-with-files

A planning skill that keeps multi-step work in Markdown files — plans, progress, and handoffs — so an agent can survive context resets without losing the thread.

25.5k 2.2k
planningcontextfilesworkflow
nidhinjsProductivity

prompt-master

A ready-to-clone library of reusable prompts and prompt patterns that installs directly into your Claude skills folder.

10.5k 1.3k
promptslibrarytemplatesbeginner

Maintain this repo? Submit an update.

Submit an update