caveman: shorter agent replies with every code span left intact
Agents burn tokens on polite openers and recaps of what they just did. caveman imposes a strict voice: answer first, then the reason and the next step, one idea per sentence. Code, commands, paths and error text stay untouched, and words like not, never and only are never dropped. The mode lasts the whole session until you say stop caveman.
Every reply follows thing, action, reason, next step, with no greeting, hedging, recap or closing line.
Built on ASD-STE100, the controlled English of aircraft manuals: 20 words a sentence at most, active voice.
Drops back to full sentences for security warnings, irreversible actions, confused users and anything persisted, such as commits or docs.
Harder levels: /caveman ultra and /caveman wenyan hand over to ultracave and megacave, and /caveman status reports the mode.
Replies in the user's language and never rewrites your prompt; only the style of the answer is compressed.
When to use it
You pay for output tokens and are tired of scrolling past pleasantries to reach a one-line fix.
The agent narrates between tool calls when one line per phase and one result line would do.
You live in a terminal and want the answer on the first line, with reasoning only when it matters.
Ask your agent
/caveman why does this React component re-render on every keystroke?
Turn on caveman and tell me why the database migration fails
stop caveman, now explain the architecture in full for a new teammate
Questions
How do I install caveman without the proxy?
Run npx skills add JuliusBrussee/caveman -g for the skill alone, or in Claude Code use /plugin marketplace add JuliusBrussee/caveman then /plugin install caveman@caveman. Type /caveman if it does not start by itself.
Does caveman work in Codex?
Yes. In Codex you call it as $caveman, and saying stop caveman switches it off.
Is caveman better than just asking the model to be concise?
Only a little: on the author's ten-question test /caveman saved about 3% more than a plain conciseness instruction at the median, and /ultracave about 35% more.
