Claude’s default output style is verbose. Every response opens with a restatement of the problem, explains what it’s about to do, does it, then summarises what it just did. In a short session that’s mildly annoying. In a long one, it actively burns context budget on tokens that carry no information.
The caveman skill enforces a compressed output mode. Claude keeps technical content — code, commands, file paths, error analysis — and drops everything else.
What gets stripped
- Preamble (“Sure, I can help with that…”)
- Transition phrases (“Now let’s move on to…”)
- Restatement of the request before answering it
- Closing summaries of what just happened
- Hedging language (“This should work, but…”)
What stays
- Code blocks, unchanged
- Commands and their flags
- Error messages and their diagnosis
- File paths and line numbers
- Opinionated conclusions (“use X, not Y, because Z”)
Why this matters at the session level
Context windows are finite. A session doing non-trivial work — reading files, running commands, iterating on output — accumulates tokens fast. Claude’s filler doesn’t just waste output budget; it also gets fed back as input on the next turn, compounding the cost.
/caveman applied at session start compounds across every subsequent response. A session that would hit the wall at turn 40 survives to turn 160.
Usage
/cavemanNo arguments. Run it once at the start of a session. The mode persists until the session ends or is explicitly lifted.
It can also be composed:
/caveman
/diagnose missing records in orders table after 2026-05-01The diagnostic output will be compressed — steps listed, not narrated.
When not to use it
Explanation tasks. If the goal is to understand why something works, the connective tissue in Claude’s output is part of the answer. caveman is suited to execution sessions, not learning sessions.