What problem does it solve? Testing systems that involve LLMs is hard in two ways: AI-generated test cases can silently lock in existing bugs as expected values, and nondeterministic model outputs cannot be verified with fixed-value assertions. This Skill provides structured procedures for both situations. ## Core Features & Use Cases - AI-assisted test generation with human review gates: Use LLMs to brainstorm test perspectives (normal, boundary, error cases), then filter false positives/negatives against the specification, register surviving perspectives as T-IDs in a ledger, and promote generalizable ones to property-based or combinatorial tests. - False-positive reduction for AI review: Apply negative prompts, severity thresholds, and formal-model translation to keep AI-generated review findings trustworthy. - Layered quality design for nondeterministic outputs: Split the system into deterministic entry/exit layers (tested with equivalence partitioning and boundary values) and a probabilistic core constrained by schema enforcement, metamorphic invariants, and multi-sampling consensus, with calibrated numeric thresholds and deterministic verdict oracles. - Use Case: When shipping a feature that embeds an LLM summarizer, use this Skill to separate prompt construction into a pure function, enforce JSON schema on outputs, define metamorphic properties, and set a calibrated similarity threshold instead of snapshot-testing flaky outputs. ## Quick Start Ask the AI to design a test strategy for a feature whose output comes from an LLM, separating deterministic layers from the probabilistic core.
