the caveman skill has been going crazy, so I built a claude-code harness to implement it on real tasks result: 26.8 percent token reduction. zero observed degradation. zero failed parses. zero verifier regressions across the test suite. (credit to @juliusbrussee)