harness-metrics-audit

To measure if your Claude Code harness is improving, you must define real success targets and compare them against visible proxy metrics. This skill interviews your goals and exposes when easy-to-count metrics drift from the outcomes you care about.

Proxy-versus-target analysis compares visible metrics like skill count or token savings against the real outcome they represent. It detects when your automation optimizes the wrong outcome by exposing the gap between the proxy and the actual goal.

You set up regression checks by designing concrete evaluations for real workflows, friction counters, and trust checks. This skill produces a small regression suite to test whether configuration changes genuinely help.

Automation metrics improve while user satisfaction drops because proxy metrics drift from real outcomes. When you optimize easy-to-count metrics like automation coverage, you may be optimizing the wrong outcome without realizing it.

Yes, you can audit Claude Code configuration changes without external dependencies. This skill requires no dependencies to interview your goals, identify misleading proxies, and produce a review cadence to detect drift.

You need a harness metrics audit when automation or speed may be optimizing the wrong outcome. It applies to harness audits, configuration changes, hooks, skills, and regressions to detect drift before users feel it.