caveman-skill
A skill that makes an agent answer without filler: fewer output tokens while code, paths and error messages stay exact.
- 81 ★ on the repository
- Last commit
- MIT
Minimal access
Why this level
- The skill holds only response-style instructions and runs nothing.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/caveman-skillDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The skill holds only response-style instructions and runs nothing.
This is third-party code. Review the repository files before installing.
What it does
caveman-skill gives an agent ten brevity rules: no greetings or "happy to help", action first and words second, no restating the question, no tool-call announcements, no closing offers of more help, and an error turns straight into a fix. Explanations come only when needed. Code, error text, file paths, technical terms, numbers and versions stay untouched. The skill trims output tokens only and leaves model reasoning alone. The README gives tiktoken `cl100k_base` measurements on four tasks: 158 tokens on average against 63, and an estimated real-session saving of 12–24%.
Who it is for. People who pay for tokens or are tired of scrolling long Claude Code and Cursor replies where half the text recaps what was already done.
Good fit when
- Before every step the agent announces it will now read a file, and you want it to simply read.
- You need short replies in a long session where closing paragraphs repeat the diff.
- The same brevity rules must go to several agents at once: ready rules files exist for Cursor, Windsurf and Copilot.
- You want the exact error text rather than a paraphrase with apologies.
Not a fit when
- You want detailed explanations and tutorial-style walkthroughs from the agent.
- Input tokens make up most of your bill: the skill does not reduce them.
- You need savings on model reasoning rather than the visible reply.
Example request
Find the TypeScript test files in the project and list the paths with no intro and no closing paragraph.Limitations
The skill cuts output text only: input and reasoning tokens do not change, so the author estimate for a real session is more modest than the measurements, 12–24% against 61% on short tasks. Style is set by instructions, and the model can slip into a long answer, especially in complex explanations. Copilot, Cursor and Windsurf use rules files rather than full skills, which must be appended by hand to `.github/copilot-instructions.md`, `.cursorrules` or `.windsurfrules`. The README uses a jokey caveman tone, while the skill itself is written plainly.
How to disable. Run `npx skills remove caveman` or delete the `.claude/skills/caveman` folder. Remove lines appended to rules files by hand.
FAQ
How do I install caveman-skill?
Run `npx skills add Shawnchee/caveman-skill`. The `-g` flag installs for all projects, `-a claude-code`, `-a cursor` or `-a windsurf` picks an agent, and `-y` removes prompts in CI.
What does the skill leave alone?
Code blocks, error text, file paths, terms, numbers and versions. Only wordy wrappers are shortened.
How do I install it manually?
Clone the repository and copy the `caveman` folder into the project `.claude/skills/` or into `~/.claude/skills/`. For Cursor and Codex the paths are `.cursor/skills/` and `.codex/skills/`.
Is the saving really 61 percent?
That is the average over four short tasks. For a real session the author estimates 12–24%, because input tokens do not change.
Which agents are supported?
Claude Code, Cursor, Windsurf, Copilot, Codex, Gemini CLI and 40+ others. Cursor, Windsurf and Copilot have separate rules files, the rest get the skill.
Discussion
No questions or reviews yet.
Related
SkillsSkill that reads your prompt history, finds the asks you repeat and turns each into a short command such as /sent
Four skills for a person with several AI assistants: a task register, evidence-based acceptance, beginner instructions
The Taixu five-turns skill pack: ideas pass divergence, convergence, verification, critique and a rescue of rejected options.
Four Claude Code skills for chess post-mortems: Stockfish-checked explanations, annotated PGN, an HTML board and a narrated video
/bro
bro
The /bro command re-explains the assistant's last reply in plain language while commands and numbers stay as they were.
A desktop app and skm CLI: one copy of each skill in a central hub, symlinked into every AI tool
README in short
A ten-rule brevity skill with token measurements and ready rules files for Cursor, Windsurf, Copilot and Antigravity. Install through skills.sh or by copying the folder.