What problem does it solve? AI products fail in ways classic UI never does — hallucination, sycophancy, mic death, TTS failure, context overflow — and most teams ship only a happy path, leaving users facing raw errors, dead spinners, or fabricated answers. This Skill provides a systematic method to enumerate every failure mode, prevent what can be prevented structurally, and design named, recoverable states for everything else. ## Core Features & Use Cases - Failure census and failure table: Walk every system layer (STT, routing, data, TTS, live components, environment) and produce a table of named cause, designed state, recovery path, and who is notified for each residual failure. - Uncertainty and trust surfaces: Apply the three-band confidence routing rule, provenance on every displayed fact, runner-up recommendations, and explainability calibrated to stakes. - Refusal and sycophancy design: Split misses into guide, honest fallback, refusal, and blip states; write deterministic verbatim copy for high-stakes refusals; counter sycophancy with dissent affordances and session resets. - Use Case: Before shipping a voice shopping assistant, run the failure census to spec what happens when the mic permission is denied, the router times out, a medicine question triggers a safety refusal, and the TTS engine dies mid-sentence — each with honest copy and a real recovery door. ## Quick Start Use the design-ai-trust-and-failure-states skill to run a failure census and design the error, fallback, and refusal states for my AI chat feature before launch.
