What problem does it solve? AI features often ship with no testable specification, leaving teams unable to judge whether LLM, RAG, agent, or classifier behavior is safe to release. This Skill turns vague AI behavior into a concrete eval plan with failure-mode rubrics, severity scoring, and mechanical ship gates. ## Core Features & Use Cases - AI-SPEC and failure-mode rubric: Defines the AI job in one paragraph, then maps failure modes with severity, likelihood, detectability, owner, and gate. - Eval case design and coverage audit: Produces concrete pass/fail cases with inputs, expected and forbidden traits, and scores coverage and infrastructure (tooling, dataset, CI, guardrails, tracing). - Mechanical ship verdict: Returns ship, ship-with-caveats, or hold based on whether severity 4-5 failure modes have cases, gates, and evidence. - Use Case: Before shipping a support-ticket triage classifier, run this Skill to write the AI-SPEC, build adversarial and regression eval cases, and get a hold verdict until the severity-5 privacy boundary is covered. ## Quick Start Use the suede-ai-eval skill to audit this AI-powered feature, define its failure modes, and design eval cases with acceptance gates.
