Skills

cave: Put Claude in a Token-Saving Caveman Answer Mode

Use this Skill when you want to save output tokens and stretch your conversation quota in Claude: after installing, type caveman mode to switch answers to a dense, no-filler style—but note it defaults to full intensity and changes how Claude writes, so it's not for output requiring formal phrasing.

Official source · asauntung/claude-skill-cave

Install

git clone https://github.com/asauntung/claude-skill-cave.git && cp -r claude-skill-cave/cave ~/.claude/skills/

Who it's for

  • Users who often have long conversations in Claude and burn through their output token quota quickly: this Skill makes Claude answer in caveman style—dense, no pleasantries, no filler—with the official goal of saving output tokens and extending usable conversation length.
  • People who use claude.ai, Claude Desktop, or Claude Code; officially, Skills are available on Free, Pro, Max, Team, and Enterprise plans (Team/Enterprise require an admin to enable Skills in Organization settings).
  • Not suitable for situations requiring full phrasing, polite lengthy explanations, or formal writing; also don't expect it to reduce input tokens—it affects the answering style. It has three intensity levels: lite/full/ultra (default full), and can be turned off with stop caveman or normal mode.
  • The official docs don't state which answer languages it supports or whether it affects accuracy beyond code-type tasks; these are unknowns you'll need to test yourself.

First try in 3 minutes

  • Prerequisite: Claude Code installed (this is the officially listed direct install path). Installation on claude.ai/Desktop works differently: you need to ZIP the cave/ folder itself, enable "Code execution and file creation" in Settings > Capabilities, then upload cave.zip via Customize > Skills and enable it—mentioned only as a secondary note, not covered in detail.
  • Install location (Claude Code): copy the cave folder from the official repository to ~/.claude/skills/:
git clone https://github.com/asauntung/claude-skill-cave.git
cp -r claude-skill-cave/cave ~/.claude/skills/
  • Trigger (low-risk, pure style switch): in the same session, type caveman mode, cave mode, or /cave. Officially, the mode takes effect immediately without confirmation.
  • Success signal: Claude's subsequent answers clearly become short, dense, and free of pleasantries in caveman style. To revert, type stop caveman or normal mode.
  • Observable safety feature: when encountering safety warnings, irreversible operations, or error-prone steps, Claude automatically exits caveman style to give full explanations, then returns to cave mode afterward; facts, numbers, code, and error text are preserved verbatim and never compressed or rewritten.
  • The official evidence doesn't state whether the Skill reads/writes files or executes commands; for your first try, stick to Q&A tasks and just observe the style change.

Use cases

  • Case 1: Save quota on long Q&A. After installing with the commands above in a Claude Code session, type cave mode, then ask multiple technical questions as usual. Expected result: answers become dense short sentences with no opening pleasantries—the official goal is that the same conversation quota lasts longer; numbers, code, and error messages still appear verbatim.
  • Case 2: Safety verification before irreversible operations. Keep cave mode on and have Claude help with a task involving deletion or deployment steps. Expected result: by official design, when reaching such irreversible or error-prone steps it automatically exits caveman style and gives warnings and explanations in full, normal language, then returns to cave mode afterward—you can confirm this behavior with a dummy task before actually executing anything. The evidence doesn't state which command types are involved; this detail needs your own verification.

Usage boundaries

  • Install paths differ by platform and must be kept separate: Claude Code uses git clone + copying to ~/.claude/skills/; claude.ai and Claude Desktop can only enable code execution and file creation via Settings > Capabilities, then upload cave.zip via Customize > Skills—there is no command-line install.
  • The mode takes effect without confirmation; it changes conversational behavior, not files. The official evidence doesn't state whether the Skill itself reads/writes files, executes commands, or accesses the network, so its runtime boundaries can't be confirmed—before installing, it's recommended to read the repository contents to understand its implementation.
  • Official restrictions: Team/Enterprise require an admin to enable Skills in Organization settings; the three intensity levels default to full, and stop caveman / normal mode can turn it off at any time.
  • Key information not yet confirmed: range of supported answer languages, whether it affects input tokens, and availability beyond the three officially listed entry points.