Decomposing LLM-Based Testing with Agent Skills: A Case Study on Numerical Inconsistencies

LLM-based testing can expose subtle reliability issues in numerical software, but monolithic prompts make it hard to inspect which guidance drives effectiveness. We study whether Agent Skills can make such workflows more explainable by factoring procedural knowledge into composable testing components. Using numerical inconsistency detection as a case study, we decompose an LLM-based differential-testing pipeline into Skills for program generation and feedback-guided mutation, then evaluate four Skill configurations. Our results show that the feedback-guided mutation Skill setting yields the largest improvement on inconsistency rate (+12 percentage points), while enriching Skills with additional procedural knowledge slightly decreases performance (−3 percentage points). An exploratory evaluation with agent-native execution reveals scalability challenges for high-iteration Skill workflows.