Caveman

Agents like to explain: a friendly opening, a recap, “I’d recommend…”. The Caveman plugin gives a connected agent the reply rules of caveman , an open-source project by Julius Brussee: drop the filler, keep every technical fact. In the author’s words, “Caveman no make brain smaller. Caveman make mouth smaller.”

The same question, from caveman’s own README:

• Without caveman: “The reason your React component is re-rendering is likely because you’re creating a new object reference on each render cycle. When you pass an inline object as a prop, React’s shallow comparison sees it as a different object every time, which triggers a re-render. I’d recommend using useMemo to memoize the object.”
• With caveman (Full): “New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo.”

How to use it

1. In the add-card menu choose Plugin…, open Caveman in the Plugins tab and press Add to canvas.
2. Pick a level on the card:
   • Lite — no filler or hedging, full sentences stay.
   • Full — the classic caveman style and the author’s default.
   • Ultra — the tersest: conjunctions go too, each fact is said once.
   • 文言文 — the same three levels in classical Chinese.
3. Draw an arrow from an agent to the card.

That’s it — the agent’s next answer already follows the rules. No restart, nothing is installed into your CLI. Change the level and the next turn gets the new one. Remove the arrow and the next turn gets a short note to answer normally again.

What stays exact

Code, commands, file paths and error messages are never shortened — only the prose around them. Security warnings and confirmations of irreversible actions come back in full sentences. Code, commit messages, documentation and pull-request text the agent writes stay in normal prose. The agent keeps answering in your language.

What the card shows

• The level, with caveman’s own example answer for it (and the same answer without caveman).
• Each connected agent: whether it works live, what it got last (the rules, a reminder, or the note to switch back) and how many turns carried the rules.
• What the rules cost: they are input for the model — the full set once per session (about 1,400 tokens), then a short reminder on every turn (about 60 tokens). Both are estimates at 4 characters per token.

Which agents it works with

It works live with Claude Code, Codex and Qwen Code — the rules go out through a hook that runs before each turn and can add text for the model. Other agents can be connected, but the card marks them Not supported: their CLIs don’t offer a way to add text to a turn from outside yet.

Dangerous mode and canvas mode don’t matter here: the plugin changes how the agent writes, not what it may do.

Where it comes from

The rules are caveman’s own — its caveman skill text and its per-turn reminder — copied into NeuroSquad unchanged from a pinned version and used under the MIT license. NeuroSquad doesn’t download anything, doesn’t send anything anywhere and doesn’t change your CLI’s settings. caveman’s separate proxy and engine are not part of this plugin.