AI Automation QA & UAT Pack

Describe any AI automation, agent, or workflow you are about to ship or accept — trigger, happy path, external dependencies, who signs off — and this skill generates a complete QA and UAT pack structured around the specific failure modes that unattended, non-deterministic systems actually hit: silent partial failures, prompt drift, idempotency breaks, and invisible errors. It operates in two modes: BUILD (generate from a description) or AUDIT (score an existing workflow's test coverage and surface the gaps). Every artifact it returns has an observable pass/fail condition — nothing vague.

A buyer building an n8n order-to-invoice automation might input: 'Webhook fires on new Shopify order, GPT-4o extracts line items and maps to billing codes, creates an invoice row in Xero, emails the customer a PDF. Peak volume 200/hour. Silent failure means a customer is never invoiced.' The skill produces acceptance criteria in Given/When/Then form, a numbered UAT script the client can run unassisted, a non-determinism test plan for the GPT-4o step, and a failure-mode matrix covering duplicate webhooks, Xero downtime, malformed payloads, and rate-limit bursts.

Sample output excerpt — Failure-Mode Matrix rows: F1 Xero API timeout | Block Xero in test env | Bounded retry x3 then alert, no silent drop | High. F2 Duplicate webhook | Replay identical order ID | Exactly one invoice created, log shows duplicate_skipped | High. F3 GPT-4o returns empty extraction | Send item-less payload | Structured rejection, no partial Xero write | High. Delivered alongside: MUST/SHOULD acceptance criteria, a regression checklist, an observability check, and a client-ready sign-off sheet.