What this checkpoint covers
Every scenario is set at Pezourd Veterinary Group, an invented group of veterinary practices with online booking and a monthly pet health plan, whose marketing lead, analyst and developer now ask an AI tool for technical work.
Illustrative example — not a real result
Pezourd Veterinary Group: the work on the table
- 6 practices, each with a page showing address, opening hours and services, and no reviews.
- The health plan system reports 760 new plans last quarter. The analyst's generated query, which joins plans to pets on the owner's ID, returns 1,140. After her fixes it returns 772, and she proposes excluding the 12 plan IDs the system does not show so the totals match.
- The tracking plan row for online booking: event
appointment_request, parameterspractice_id,speciesandappointment_type, consent category analytics. - The generated booking tag sends
vet_appointmentwithpet_name,owner_phoneandpractice, on pages matchingbook: the booking confirmation and a blog post on books for new puppy owners.
Task
- Answer all eight in writing, action and reason, before reading on.
- Score yourself against the explanations; the right action for the wrong reason scores as a miss.
- Seven or eight: move on. Fewer: reread the page linked from each miss.
Questions
- The marketing lead types "health plan conversion rate by practice last quarter, plus a tag for bookings" into the approved tool. What should exist before either request is sent?
- The analyst plans to paste rows of the plan export, with owners' names and phone numbers, into her personal AI account "so it can see the data". What does the model need instead?
- Her first query returned a neat table by practice in seconds. The marketing lead wants it in tomorrow's report. Your reply?
- Her corrected query sits a little above the plan system's total. What is wrong with her proposal?
- Read the generated booking tag against the plan row. What has to change before anything else is tested?
- The developer asked the tool for location markup and pasted the result into each practice page by hand, opening hours and a star rating included. What should he have asked for, and what comes out?
- The developer tested the booking tag once, on the confirmation page, after accepting the banner. What is missing?
- The developer wants to publish from his own workspace this evening, with no description; the analyst has saved her corrected query to her downloads folder. What should happen to each?
Answers explained
- The definitions and the plan row. The question needs every term defined, and the tag needs its tracking plan row. The AI-written code review checklist puts both before anything reaches the model, because without them there is nothing to check the output against.
- Field names, row meanings and invented rows, in an approved tool. Owners' details add nothing to a query and are Personal DataAny information relating to an identified or identifiable living person, such as a name, email address, location data or an online identifier. the task does not need, and a personal account is not an approved tool. The tutorial on minimising data before prompting starts from the task and keeps only the fields it needs.
- Running proves the syntax, not the answer. A Language ModelAn AI system trained on large amounts of text that writes by predicting which piece of text is likely to come next. writes a likely-looking query from what it was shown, and a join on the owner's ID repeats each plan once for every pet that owner has. The routine for checking AI-written SQL proves it against the definition and an independent total first.
- Excluding plan IDs makes the numbers agree without making them right. A gap gets an explanation, or the checks run again. The diagnosis for AI report numbers that do not match gives the same rule.
- Names and personal data.
vet_appointment,pet_nameandpracticeare not in the plan, andspeciesandappointment_typeare missing, so reports would split.owner_phonecomes out. The tutorial on proving AI-written tracking code checks names and values before preview: a tag can fire while sending the wrong things. - A mapping from the page template's fields, not finished markup per page, so the markup follows the opening hours when they change. The rating comes out, because the practice pages show no reviews, and structured dataStructured data vocabulary. describes only what a visitor can see. See the tutorial on checking AI-generated structured data.
- The pages that should stay quiet, and the refusal. The pattern also matches the blog post, so the trigger is tested against both lists. Preview runs again with the banner refused: Google's tags adapt to the choice by themselves, and the preview shows whether a consent requirement held the tag back. See the consent mode guide.
- Through the container's rules, and saved where others can find it. The workspace description says an AI tool drafted the tag and how it was tested, and someone other than its author reviews and publishes it, as the container governance guide sets out. The query goes into the team's shared library with its definition, checks and the PromptThe instruction and material given to an AI language model, which works like a brief: the model knows language, not the audience or product, unless told. that produced it.
Where to go next
Missed the third or fourth? Redo the query half of the sewing machine retailer exercise. The advanced AI in marketing level covers re-testing a saved prompt after a model change. The next module turns from code to customers, and its exercise is the coffee machine retailer's chat assistant.
