LLM
Lets you ask many AI models questions from the terminal and keeps a log of every answer
by Simon Willison
Caveman
by Julius BrusseeSkillWrite code
Makes your AI agent put the answer first and drop filler, with code and facts kept exact
The problem it solves
You ask your agent a quick question and get three paragraphs of preamble before the one line you needed.
For Coding agent usersPeople paying per token
- Before
- You scroll past greetings, recaps and offers to help to find the actual answer.
- After
- Replies lead with the answer in short sentences, with code, paths and numbers unchanged, until you say stop caveman.
- Stars
- ★ 116k
- License
- Apache-2.0
- Cost
- Free
- Level
- Beginner
- Contents
- Instructions only
Overview
Caveman is a skill that changes how your AI agent writes its replies. It puts the answer first, cuts greetings, hedging and recaps, and keeps code, commands, numbers and error messages exactly as they are. It switches back to full sentences for security warnings, irreversible actions or when you seem confused, and stays on until you say stop caveman. It suits people who use coding agents and want shorter replies that cost fewer tokens.
It shortens what the agent says without touching your prompts, and its author publishes the cases where the savings turn out small.
Try these first
Start with the first one: it is a small task you can check easily.
- /caveman Then answer this question I asked earlier, so I can compare the two replies: [paste question]First try
- Explain this error and the fix: [paste error message]
- /caveman status
- stop caveman
Example prompts written by Agent Flows Hub.
What you get
From the creator's documentation.
| Output | What it is | Reads |
|---|---|---|
| Answer first | Puts the answer before the reason and the next step | Agent replies |
| No ceremony | Drops greetings, hedging, recaps and closing offers | Agent replies |
| Exact payload | Keeps code, commands, paths, numbers and error text unchanged | Agent replies |
| Clear when it matters | Switches to full sentences for security warnings, irreversible actions and confused users | Agent replies |
| Stays on | Lasts the whole session until you say stop caveman or normal mode | Agent replies |
Good to know
Get shorter replies from a coding agent during a long session. The skill makes the agent answer first in short sentences and leave code, paths and error text unchanged.
Limits
- The skill's own measurement found about 3% fewer output tokens than simply asking for concise answers, within the noise; the separate ultracave skill cut more.
- If you pay per request rather than per token, shorter answers save nothing, as the project README notes.
Notes by Agent Flows Hub (original project name: caveman), written from the creator's documentation, not from running it ourselves. caveman skill: SKILL.md · caveman skill: README and measurements · Caveman project README: measurements and per-request billing note
Stuck setting it up?
Tell us what happened. We take a look first; if it needs paid work, you get an estimate before anything starts.
More to write code
Other skills
- Charlie AI CFO CFO-style thinking on cash, runway and hiring for self-funded founders
- Company Research Report Writes a full value-investing report on one listed company from four investors' views
- Competitor Ad Teardown Breaks down a competitor's public ads into hooks, offers, calls to action and test ideas
- Competitor Analysis Maps your main competitors and shows where you can stand out
- Consulting-Style Charts Turns notes and numbers into polished, consulting-style slides and reports
- Contract Risk Review Reads a contract, explains risky terms in plain English and drafts a redline