Developer Pre-Commits Pass Criteria to Git Before Testing Claude Code AI Skills

A developer built three Claude Code skills for AI product managers and, before running any evaluations, committed the pass criteria to a Git repository to prevent post-hoc goal-shifting. The first evaluation run failed, revealing a flaw in the /build-or-not skill, which confidently recommended against building a feature despite having no evidence to draw on. A single rule fix — 'no sample, no decision' — resolved the issue, and the second run passed all gates. Testing the /agent-trust-review skill took four runs, with every failure traced back to errors in the test setup rather than the model itself. Across eight test cases and three runs each, the custom skills consistently outperformed plain Claude on structured refusals and pre-defined decision criteria, at a total cost of roughly $2 per full run.