Technique 04 / 14
Caveman mode
Compress every reply to the fewest tokens that still carry the technical substance.
Interactive model
See the change.
Task: “Fix the login bug.” Compare the default reply with Caveman’s.
Fix the login bug
Fixed. Null check, auth.ts:42. Tests pass.
Conceptual demonstration. Counts illustrate the flow, not measured production savings.
How it works
Less context. Same thread of work.
Replies full of preamble, recaps and follow-up offers are billed as output and then re-read as context on every later turn. Caveman instructs the agent to answer first and cut filler, aiming for at most half its usual prose, across progress updates and final answers alike.
Code, commands, paths, numbers, errors and negations are kept exactly; Caveman removes ceremony, not facts. Security warnings, irreversible actions and ambiguous situations still get clear normal prose. Caveman is off by default and stays on for the session once enabled.
In practice
The operating loop
- 01
Answer first
The result leads; walkthroughs appear only when asked for.
- 02
Cut, then cut again
Filler, restatement and recaps are dropped before the reply is sent.
- 03
Keep the payload
Code, commands, paths, numbers and errors survive unchanged.