AI Radar

Caveman skill claims to cut coding-agent tokens by 65%

The Caveman repository is a skill and proxy for coding agents that makes the agent write terse output to save tokens. Its README claims a 65% token cut and also promotes a new agent, Caveman Code. The benchmarks are the project's own.

What to do
Install the skill and compare token use on your own tasks against the README claims.
Who should care
Developers using Claude Code, Codex or similar coding agents who pay per token.

Why it matters: Caveman is Apache-2.0, so developers can install the skill and test whether its claimed token savings hold in their own agent sessions.

Key facts

Claims 65% fewer tokens
Claims 33.2% fewer input tokens across 54 Claude Code runs
Licence
Apache-2.0
Claims Caveman Code is 21% cheaper than Codex on Terminal-Bench 4
Claims 14% cheaper than Claude Code, 97 vs 83 tasks solved

Timeline

Coverage

Evidence

Medium evidence: several community reports point the same way, but there is no official confirmation yet.

Details written Nov 14, 17:38 UTC from the sources listed above.