I built an AI front desk for the Meow Meow Veterinary Hospital competition round. You type what you need in your own words, and it books, cancels or reschedules against a live Google Calendar. Every request is logged to Google Sheets, and it sends a confirmation email. The design rule: the LLM only turns messy text into structured JSON. All the rules and writes are deterministic code. A misread message can cause a question or a refusal, but never a bad booking. The problem I hit: my form test passed, but the public form threw “Workflow could not be started” in a real browser. n8n won’t start a form workflow that also has a Respond to Webhook node. My test went through the API, which skips that check, so it was a false pass. How I solved it: 1. Split it into a form workflow and a main pipeline workflow. 2. The form workflow passes fields to the main one and shows the HTML result. 3. Re-verified it end to end, and wrote the correction into my test report instead of hiding it. Other bugs live testing caught: - The email step crashed after the booking was saved. It used a helper the n8n code sandbox doesn’t have. I made both email steps crash-proof. - “Wellness exam for a bird” booked an Exotic Exam. A prompt fix plus a rule check solved it. What I’m proud of: - 20 of 20 required tests and 14 extra cases passing live - Rollback if the sheet write fails after the calendar event is created - Replay-safe: the same request ID returns the original answer instead of double-booking - Prompt injection and medical-advice requests are refused or handed off Lesson: a test only counts if it uses the same path your users do. Screenshots are attached: the workflow canvas, the form, and the records sheet. Happy to share the structure or prompts if anyone wants them. How do you test your n8n workflows? 👇