Make each drink once while saying the measurements aloud, turn the transcript into a fixed-format recipe card with AI, then ask AI for a ten-question quiz that uses only those cards. You supply every number; AI supplies the layout and the questions. Check each card against a real cup before any new hire trains from it.
Online courses and certificates teach bubble tea in general. A new hire at your counter has to learn your cups, your syrup pumps, your sweetness ladder and your pearl rules, and only you can supply those. That is why the AI's job here is narrow: format what you say, then test whether staff remember it. Left alone, it fills gaps with "standard" measures that are not yours, and a confident wrong number on a laminated card is worse than no card.
The examples use a one-counter shop with a 12-drink menu, two cup sizes and a four-hour rule for cooked pearls. The figures are invented for illustration. The milk and pearl amounts match the ones in the bubble tea inventory forecast sheet, because the same card quantities feed both jobs.
Record each drink once, out loud, at the counter
The most accurate source of measures is you, making the drink at the station. Prop up a phone, make each drink at normal speed and narrate as you go. Five minutes per drink is enough, and the mess is useful: the stray remarks about what trips people up are exactly what a good card needs. Then turn the recording into text. Dictation in a notes app works, so does any transcription tool, and the transcript becomes the source of truth. The AI never sees your video, only the words, so anything you do silently will not reach the card.
Narrate in a fixed order so every transcript has the same bones:
- The drink's name, exactly as the menu and the till call it.
- The cup size and where the fill line sits.
- Every ingredient with its measuring tool and amount: "scoop A", "four presses", "the 200 line on the jug", never "a bit".
- The order you build in.
- What changes for sweetness and ice.
- The shake count and the seal check.
- Anything made in advance, with the hold time from your supplier's instructions.
- The one mistake new staff always make.
If you would rather record a video and work from that, the tutorial on creating procedures from screen recordings with AI covers the same transcript-first habit for other jobs in the business.
Define the sweetness and ice ladders before writing any card
Most shops offer the same choices on every drink: a sweetness scale and an ice scale. Define each once and let every card refer to it. Then a change to the ladder is a one-page edit, and no card can quietly drift away from the others. Here is an illustrative ladder for syrup-sweetened drinks in the large (700 ml) cup, with a pump that delivers 10 ml per full press.
| Sweetness | Syrup in a large cup | Presses |
|---|---|---|
| 100% | 40 ml | 4 |
| 75% | 30 ml | 3 |
| 50% | 20 ml | 2 |
| 25% | 10 ml | 1 |
| 0% | none | 0 |
The ice ladder works the same way, and it carries one detail that trips people up: when ice comes out, the shop replaces some of it with tea so the customer does not feel short-changed. In the example, regular ice is one scoop; less ice is half a scoop plus 30 ml of extra tea; no ice is none plus 60 ml of extra tea. Your amounts come from testing, and some shops simply fill to a lower line instead.
With both ladders in place, an order becomes arithmetic anyone can follow. Take a ticket reading "large classic milk tea, 75% sweet, no ice". The card and ladders give 70 g of pearls, 3 presses of syrup (30 ml), no ice, 260 ml of tea (the 200 ml on the card plus 60 ml for no ice) and 110 ml of milk. A new hire should be able to get from the ticket to those five numbers without asking anyone.
Two cautions. Brown sugar drinks and fruit teas often sweeten differently (syrup lining the cup, sweetened pulp), so give them their own ladders rather than forcing one scale across the menu. And give the regular 500 ml cup its own column, then test it by tasting, because large-cup amounts do not scale cleanly down. When you paste the ladders into the AI, say that they are fixed. Otherwise it may "tidy" your 40 ml into its own idea of a full-sweet drink.
The card prompt, a sample card and what the AI gets wrong
Give the AI the layout, the rules and the transcript, and tell it what to do when something is missing. The instruction to write MISSING instead of guessing does most of the work.
Turn the transcript below into a recipe card for new staff. Use this exact layout: Name; Cup; Build order (numbered, one action per line, each with amount and tool); Sweetness (refer to Ladder A, do not restate numbers); Ice (refer to the Ice ladder); Toppings; Allergens; Hold times; Mistakes to avoid; Version and date.
Rules: use only what is in the transcript or the ladders I attach. Do not add steps, brands, times, temperatures or garnish. Keep my units (ml, g, presses, scoops). If a field is missing, write MISSING and name what I must supply. Do not guess allergens.
[paste the ladders]
[paste the transcript]Here is why the MISSING rule matters. Imagine this fragment of a transcript: "you grab a good scoop of pearls, in the cup first, then for full sugar I do four pumps, then the tea, about two hundred from the brewer, then milk to the line." A careful card writes the pearls as "MISSING: weight of a good scoop", the syrup as "Ladder A (100% = 4 presses)", the tea as "200 ml from the brewer" and the milk as "MISSING: amount, and which line". Two gaps flagged in one sentence is a good result. You answer them (scoop A is 70 g, the milk is 110 ml) and the card has numbers where there was a gesture.
A first pass for the classic milk tea comes back like this (illustrative):
| Field | Card content |
|---|---|
| Name and cup | Classic milk tea, large (700 ml) |
| Build order | 1. 70 g pearls (scoop A) into the cup. 2. Syrup per Ladder A. 3. Ice per the Ice ladder. 4. 200 ml black tea from the brewer. 5. 110 ml milk. 6. Seal, shake 6 times, check the fill line, add a wide straw. |
| Sweetness | Ladder A. Ask every time; never assume. |
| Ice | Ice ladder. Replace removed ice with tea. |
| Allergens | MISSING: owner to complete from ingredient labels (milk is used) |
| Hold times | Pearls: bin 4 hours after cooking (shop rule). Brewed tea: MISSING |
| Mistake to avoid | Sealing before checking the fill line |
| Version | v1, [date] |
That is a good draft, and the first pass of a real run usually has a few slips as well. These are the ones to look for, with the fix for each:
- It wrote "steep the tea for 5 minutes". You said "tea from the brewer". Delete anything that is not in the transcript.
- It converted 200 ml to 6.8 fl oz. Restore your units; staff measure in millilitres.
- It wrote "shake well" where you said six shakes. A count is checkable and "well" is not, so put the number back.
- It added "garnish with a mint leaf", which nobody said. Remove it.
- It filled the allergen line with "contains dairy" on an earlier attempt. That may be right, but a guess is not a source (the next section covers why).
Allergens and hold times: lines the AI must never fill in
An AI model cannot know what is in your syrup, creamer or topping, and a wrong allergen line is a safety problem, not a style problem. Keep one allergen sheet for the shop: ingredient, brand, what it contains, and the date someone read the label. Each card's allergen line is the sum of the sheet's lines for the ingredients in that drink. The AI can help lay out the sheet and cross-check that every card ingredient appears on it, but a person reads the labels, and the check is repeated whenever a supplier changes a product.
| Ingredient | Brand and pack | Contains (copied from the label) | Label read on | Read by |
|---|---|---|---|---|
| Fresh milk | [brand], 2 L carton | Milk | [date] | [initials] |
| Non-dairy creamer | [brand] | Copy from the label; some creamers sold as non-dairy can still contain a milk protein | [date] | [initials] |
| Tapioca pearls | [brand] | Copy from the label | [date] | [initials] |
| Syrup for Ladder A | [brand] | Copy from the label | [date] | [initials] |
The same applies to hold times. The four-hour pearl rule in the example is the shop's decision, set from the supplier's instructions and local food-safety guidance; another shop's rule may be shorter. Put the rule on the card and a "cooked at" label on the tub, and never let the assistant recall a number from general knowledge. If customers ask allergy questions through a chatbot or on the phone, whether an AI chatbot should answer allergen questions sets out why those answers need a person and a verified sheet, and the quiz below tests that habit directly.
Check the card works with someone who has never seen the recipe
A card that reads well to the person who wrote it can still fail at the counter. Run five checks before it is printed.
- Make the drink from the card alone, with a measuring jug, and compare it with your usual drink for taste and fill line.
- Weigh one scoop of pearls on a kitchen scale. It should read 70 g; if it does not, fix the scoop or the card.
- Taste the 25%, 50% and 100% levels side by side. Ladders that look sensible on paper sometimes taste wrong.
- Hand the card to someone who has never seen the recipe and watch without helping. Every question they ask is a missing line.
- Compare the card's milk and pearl amounts with your stock sheet, so ordering and training use the same numbers.
Here is how the fourth check pays for itself (an illustration). A first tester followed the card for a 50% drink and pressed the pump four times, because Ladder A sat on a separate page and the card only said "per Ladder A". The result was a drink at 100%. The fix was not to retrain the tester; it was to print a small ladder strip down the edge of every card. If a new hire has to ask, fix the card, not the person.
A ten-question quiz from your cards, in a form that marks itself
Once three or four cards have passed the checks, ask for a quiz. The rule that makes it trustworthy is that every question must quote the card line it came from. That turns your audit into ten quick ticks, and it stops the assistant sneaking in "common knowledge" that is not your recipe.
Using ONLY the recipe cards and ladders below, write a 10-question quiz for a new bubble tea hire: 5 multiple choice (four options), 3 scenario questions with four options, 2 short answer. Include 2 must-pass questions, one on allergens and one on pearl hold time. For each question give the answer and quote the line of the card or ladder it comes from. If the cards do not contain the answer, do not write the question. No trick questions. Plain English.
[paste the cards and ladders]A usable result for the sample card looks like this (illustrative, with answers):
- Multiple choice. A large classic milk tea at 50% sweet needs how many syrup presses? 1, 2, 3 or 4. Answer: 2 (Ladder A, 20 ml).
- Multiple choice. What goes into the cup first? Ice, tea, pearls or milk. Answer: pearls (build order, step 1).
- Multiple choice. "Less ice" means: no ice; half a scoop plus 30 ml extra tea; one scoop; two scoops. Answer: half a scoop plus 30 ml extra tea (Ice ladder).
- Multiple choice. Put the build in the right order. Four candidate orders are offered. Answer: pearls, syrup, ice, tea, milk, seal and shake.
- Multiple choice. How much milk goes into a large classic milk tea? 90, 100, 110 or 120 ml. Answer: 110 ml.
- Scenario. A customer wants no ice. What do you add? Answer: 60 ml extra tea (Ice ladder).
- Scenario, must-pass. The pearls in the tub were cooked at 12:10 pm and it is 4:30 pm. May you serve from the tub? Answer: no; the four-hour hold ended at 4:10 pm, so bin them and cook fresh.
- Scenario, must-pass. A customer says she is allergic to dairy and asks whether the milk tea is safe. Answer: do not guess; check the allergen sheet and the labels, ask the manager, and if it cannot be confirmed say the shop cannot guarantee it.
- Short answer. How many millilitres of syrup go into a large drink at 75% sweet? Answer: 30 ml (3 presses).
- Short answer. You knock a finished drink off the counter. What do you do? Answer: remake it from the card and write the cup, drink and reason on the waste log.
The last answer links the quiz to the stock sheet: the waste log is what tells you whether the 4% cup allowance in the inventory forecast is realistic. A sensible rule is to pass at 9 out of 10 with both must-pass questions correct. Anyone below that rereads the cards and retakes a fresh version, which you can ask the AI to write from the same cards.
To make the quiz mark itself, build it in a free form tool. In Google Forms, open Settings and turn on "Make this a quiz", then use "Answer key" to mark the right answers and set points; Google's help page on quizzes describes the options. Answer keys work with short answer, multiple choice, checkboxes, dropdown and grid questions, so an ordering question becomes a multiple-choice with four candidate orders, as above, and the two short answers are easiest to mark by hand. You can release grades straight after each submission, and the Responses summary lists the questions people miss most, which is the most useful output of the whole exercise.
The first week for a new hire, built around the cards
Cards and a quiz do not replace shadowing; they give shadowing a spine. One week might run like this.
| Day | Learn | Practise | Check |
|---|---|---|---|
| 1 | The station, both ladders, cards 1 to 3 | Each drink twice with the card in hand | Owner tastes and checks fill lines |
| 2 | Cards 4 to 6 and the topping scoops | Live orders beside a trainer | Trainer logs remakes |
| 3 | Cards 7 to 12, pearl cooking, hold times, allergen sheet | Cook and label one pearl batch | Quiz, first attempt with cards open |
| 4 | Till, special requests, the waste log | Run the bar with the trainer on the till | Remakes for the shift |
| 5 | Revision of any missed questions | Three random drinks from the card stack, no prompts | Quiz with cards closed: 9 of 10 plus both must-pass |
Keep the sign-off in one place so you can see who has covered what. The tutorial on building a staff training matrix with AI shows a simple version, and the fast food crew reference cards tutorial is useful if your shop also sells snacks or hot food with their own stations.
Keeping cards current when the menu changes
Every card carries a version and a date. When a recipe changes, edit the card by hand first, then use the AI for the follow-through, which is where errors creep in. Paste version 1 and version 2 of the card and ask: "List every difference as a bullet. Then list every quiz question whose answer is now wrong."
A realistic change shows why. Say you trim the milk in every milk tea from 110 ml to 100 ml. Six cards change, and quiz question 5 has a new answer. On the stock side, the weekly milk figure drops from 98.6 litres to 89.6 (896 milk cups x 0.100), about 9 litres a week less to buy. If you change the card and forget the quiz, new hires are tested on the old amount; if you change both but not the forecast sheet, you order milk you will not use. One edit, three places, and the AI's job is to list them.
If some staff read another language, translate the finished card, not the transcript, and ask a bilingual colleague to check every number and every food-safety line. Translation tools can drop a unit or turn "press" into a different action. The tutorial on restaurant menu translation with AI covers the checks in more detail.
What repeated quiz misses say about your cards
The quiz is a diagnostic as much as a test. Say six new hires took it over a year, and question 6 (no ice) was missed by four of them. That is not four careless people; it is a card that buries the "add 60 ml of tea" rule. You move it to the top line of the card and the ladder strip, and the next six hires miss it once. When the same question catches several people, change the card.
Track three other numbers for each new hire: remakes in the first week, the day they made their first unaided correct drink, and whether the fill line was right on a spot check. Compare each hire with the last. No benchmark from outside your shop is needed, because the trend within your own counter is the only one that matters.
The time sum is easy to make. Building 12 cards takes roughly five hours, counting about 25 minutes a drink to record, format and check, plus an hour for the quiz. If you spend six hours explaining the menu to each new hire now and the cards halve that, four hires a year give back twelve hours for a one-off cost of about six. That figure is illustrative, so use your own hiring pattern, but the shape holds: the work is front-loaded and the benefit repeats. None of this needs a paid AI plan. What costs time is the checking, which is the part you should not hand over.
Questions owners ask about AI-made recipe cards and quizzes
Should the cards carry a photo of the finished drink?
Yes, but take it yourself. A photo from your own counter, in your own cup at the correct fill line, shows staff what good looks like. An AI-generated picture sets a target your shop cannot reproduce, and it makes remakes and complaints more likely because staff compare real drinks with an image that was never made.
Is it safe to paste my signature recipes into ChatGPT or Claude?
Treat a recipe like any other business secret. On consumer plans, switch off the model-training setting in privacy settings; business plans do not train on your content by default. If a signature ingredient is truly secret, leave it out of the transcript, let the AI format the rest, and type that line into the finished card by hand.
How often should staff retake the quiz?
Set it after any menu change that touches their cards, and again about three months after they start, when habits have formed. Use a fresh version generated from the same cards so staff are tested on the recipe rather than on remembering last time's answers. Keep the allergen and hold-time questions in every version.
What if a team member finds reading long cards hard?
Keep one action per line, use numerals and units only, and add a photo of each measuring tool next to its amount. Walk through the first cards aloud with a trainer, and let that person answer the quiz verbally. The aim is to prove they can make the drink correctly, not that they can read a paragraph.
Further reads
- How to Build a Kitchen Training Manual With AI — A fuller training manual for a kitchen with more stations.
- AI Employee Onboarding: Plan a New Starter's First Two Weeks — Plan a new starter's first two weeks beyond the recipe cards.
- Recipe Costing Spreadsheet for Restaurants: Build It With ChatGPT — Turn card quantities into cost per drink and margin.
- How to Write SOPs With AI: From Rough Notes to Clear Procedures — Write opening, closing and cleaning procedures the same way.
- Restaurant SOP Template: Opening and Closing Checklists With AI — Opening and closing checklists for a counter-service shop.
- AI Menu Making Tool: Allergen Icons and Dietary Labels Before You Print — Allergen labels and menu icons before anything is printed.
- Coffee Shop Pricing: Cost per Cup and Margins Worked Out With AI — Price drinks from the per-cup amounts on your cards.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Google Docs Editors Help on making a Google Form a quiz (settings, answer keys, grade release); OpenAI and Anthropic privacy settings help pages. Hold times for cooked pearls come from your own supplier's instructions and local food-safety rules. All shop figures in the examples are illustrative.