What problem does it solve? Probabilistic AI and LLM features cannot be declared done with a simple "all tests pass" checklist, because the same prompt produces varying outputs across users, sessions, and model updates. This Skill defines a five-layer definition of done so teams know exactly when an AI feature is calibrated and safe to ship. ## Core Features & Use Cases - Five-layer done bar: Checks deterministic floor, distributional quality bands, failure-triage playbook, tripwires with rehearsed rollback, and a post-ship eval loop, refusing to report done on a partial bar. - Thin orchestration: Delegates deterministic checklists to /success-criteria and eval suites to /eval-harness instead of duplicating them, then fills the three gaps neither covers. - Calibration doc generation: Emits a one-page per-feature calibration document from a template, with deterministic path selection and a green/red readiness summary. - Use Case: Before shipping an LLM-powered support chatbot, run this Skill to confirm eval bands pass, a kill switch has been rehearsed with real evidence, and production outputs feed back into the eval suite. ## Quick Start Ask the agent to run the ai-done readiness check on your AI feature to verify all five layers are green before shipping.
