The AI prompts that help bakers most give the assistant your real costs and facts first, then ask for one job: a cost-per-item breakdown with price options, three caption drafts in your voice, or a reply to an order, complaint or allergen question that you check before sending. Twelve such prompts follow, ready to copy.
They work in ChatGPT, Claude or Gemini. Two rules apply throughout: check every number in a spreadsheet, because chat assistants make arithmetic slips, and never let AI be the source of allergen information. The prompt can phrase the answer; your recipe records must supply it.
Paste this bakery fact sheet before any prompt
Generic prompts get generic answers. The single biggest improvement is a short fact sheet you paste at the top of a conversation, or save once in a ChatGPT or Claude project or a Gemini Gem (becoming a skill from November 2026). Fill it in once; update it when prices or hours change.
BAKERY FACT SHEET
Name: [bakery name] Type: [e.g. neighbourhood bakery with a small café counter]
What we sell: [5-10 best sellers with prices]
Ordering: [how custom orders work, notice needed, deposit, collection/delivery]
Opening hours: [days and times]
Tone: [e.g. warm, plain, a little dry; no exclamation marks]
Never say: [e.g. "artisanal", "to die for", "guilt-free", "best in town"]
Allergen policy: [e.g. "We handle nuts, gluten, dairy, egg and sesame in one kitchen.
We can't guarantee any item is free from these. Ask staff for the allergen folder."]
Complaints: [what you offer: replacement, refund, credit; who decides]
Keep the sheet to one screen. A page of brand philosophy dilutes the facts the assistant actually needs, such as your notice period for custom orders or the fact that you don't deliver on Sundays. If you want to go further, add two or three captions or replies you've written yourself and liked, labelled "examples of our voice". Assistants copy examples far more faithfully than they follow adjectives such as "friendly" or "authentic".
The "never say" line does a lot of work. AI copy for food businesses drifts towards the same handful of adjectives, and banning your least favourite ones makes the rest sound more like a person wrote them.
Pricing prompts
1. Cost one item properly
Use it when you're pricing a new bake or haven't recosted since your flour and butter prices went up. The prompt asks the assistant to lay out its working, so you can check each line.
Here is my recipe for [item], which makes [yield] pieces:
[ingredient, quantity used, pack size, pack price] (one per line)
Labour: [minutes of hands-on time per batch] at [hourly wage].
Packaging per piece: [cost]. Oven/energy per batch: [estimate].
Calculate cost per piece, showing each line: ingredient cost used,
labour per piece, packaging, energy. Then total.
Use code to calculate if you can, and list any assumption you made.
Check: recalculate one ingredient line yourself. The usual slips are unit mix-ups (grams against kilogram pack prices) and dividing by the wrong yield. Say a sourdough batch of 12 loaves uses $13.20 of ingredients, 72 minutes of hands-on work at $15 an hour ($18) and $4.80 of energy and bags; that's $36 or $3.00 a loaf. If the AI's figure is far off, ask it to show the division it used.
2. Turn a cost into price options
My [item] costs [cost per piece] to make (from the costing above).
I aim for [e.g. 65-70%] gross margin on counter items.
Give me three price points: one at my target margin, one 10% below,
one 10% above. For each, show gross profit per piece and per week
if I sell [weekly quantity]. Don't tell me what competitors charge;
you don't have that data.
Here's what a sensible answer looks like, so you can spot a bad one. Take the $3.00 loaf and a 65% margin target: the price that hits it is $3.00 ÷ (1 − 0.65), about $8.57, which you'd round to $8.50 for a gross profit of $5.50 (about 65%). Ten per cent lower, rounded to $7.75, gives $4.75 profit (about 61%); ten per cent higher, rounded to $9.40, gives $6.40 (about 68%). At 120 loaves a week, the gap between the lowest and highest option is about $198 of gross profit a week. That's the number to think about, and it's easy to miss if you only look at the price on the shelf.
The last line of the prompt matters. Chat assistants will happily invent "typical local prices" if you let them. What the market will bear is something you find out by looking at nearby bakeries and watching your own sales, not something the AI knows; the limits of AI pricing research goes into why.
3. Build a custom cake quote
Using my fact sheet, draft a quote for this custom cake order:
[size/tiers, servings, flavours, decoration details, date, collection or delivery].
My pricing: base price per serving [x] for [style], plus [y] per hour
for decoration beyond [z] hours, plus delivery [rate].
Show the calculation line by line, then a short, friendly quote message
with the total, the deposit ([%]) and the date the balance is due.
Flag anything in the order I haven't priced.
"Flag anything I haven't priced" catches the sugar flowers or the extra tier the customer mentioned in passing. If custom orders make up a big share of your work, the full enquiry-to-quote setup is in taking custom cake orders with an AI enquiry assistant.
Filled in for an illustrative three-tier birthday cake (60 servings, lemon and raspberry, buttercream with fresh flowers, four hours of decoration against two included, delivery on a Saturday), the calculation that came back read:
Base: 60 servings x $4.50 $270.00
Extra decoration: 2 hours x $30 $60.00
Delivery $25.00
Total $355.00
Deposit (30%) $99.00
Balance due 7 days before $256.00
Not priced: "fresh flowers" - please confirm
Two things to catch. The flag worked: the customer's flowers were never in the pricing, so that line becomes a question for the florist or an extra charge. The deposit didn't: 30% of $355 is $106.50, and the assistant took 30% of the cake before extra decoration and delivery ($99 is 30% of $330, rounded). Decide once whether your deposit is on the whole order, write that into the prompt, and check the deposit line every time.
4. Tell regulars about a price rise
I'm raising prices on [items] from [date] by [amounts].
Reasons (true ones only): [e.g. butter up X% since spring, energy costs].
Write a short counter sign (under 50 words) and an Instagram caption
(under 80 words). Plain and honest, no apology spiral, no guilt.
Mention what hasn't changed: [e.g. same recipe, same flour, still baked here at 5am].
Given "butter up 18% since spring, energy costs" as the reasons, an illustrative counter sign came back as: "From 1 November, sourdough goes from $6.50 to $7 and croissants from $3.20 to $3.50. Butter prices have almost doubled this year, and energy has followed. Same recipes, same flour, still baked here from 5am." It reads well, and "almost doubled" is wrong: the figure you gave was 18%. Assistants round small truths up into dramatic ones when asked to persuade. Put your own figure back, or drop the number and say "butter and energy both cost us noticeably more than last year".
Check: the reasons are yours, not the AI's. It will suggest plausible ones ("rising costs across the industry") if you don't give it real ones. More on timing and wording in planning and announcing a price increase.
Caption prompts
Bakery captions have one job: make someone want to come in, today or on a specific day. That means the useful ones nearly always contain a product, a time and a reason. AI is good at producing tidy sentences and bad at knowing which of your bakes are worth talking about this week, so you choose the subject and supply the true detail; the assistant does the wording. Keep them short. On a phone screen, anything after the second line is hidden behind "more".
5. Launch a new bake
Fact sheet above. New bake: [name], [price], available [days/times].
What makes it different: [2-3 true details: e.g. brown butter, 48-hour dough].
Write 3 caption options under 60 words. One plain and descriptive,
one with a small behind-the-scenes detail, one that's a single line.
No invented ingredients, awards or "limited" claims unless I've said so.
For a brown butter cardamom bun at $3.80, on Fridays and Saturdays from 8am, three options might come back like this (illustrative):
1. New on the counter this Friday: brown butter cardamom
buns, $3.80. Soft, spiced, and gone by lunchtime.
2. Our grandmother's cardamom recipe, brought back to life
with brown butter. Fridays and Saturdays from 8am.
3. Brown butter. Cardamom. Friday, 8am.
Option 2 is the one to watch: nobody mentioned a grandmother, and invented heritage is exactly the kind of claim a regular will ask about. Option 1 has a "gone by lunchtime" line you can't promise on a slow Saturday. Option 3 is usable as it stands. A fair edit takes the first sentence of option 1 and adds a true detail from the prompt, such as how long the butter is browned, so the caption says something only your bakery could say.
Asking for three different styles, rather than three versions of one, gives you real choice. Pick one and edit a word or two; that last human edit is what stops captions reading as automated.
6. Sold out, back tomorrow
[Item] sold out at [time] today. Tomorrow we'll have [quantity/when].
Write a one or two sentence post that thanks people, says when it's back,
and suggests [alternative item] for today. Light, not boastful.
Sell-out posts are among the easiest to get wrong: "Sorry, we're sold out AGAIN!" reads as a humblebrag. The prompt steers towards useful information instead. Compare "Sorry, sold out AGAIN!" with "The almond croissants went by 10.40, thank you. Another tray is in at 7am tomorrow; for today, the pear and frangipane tart is still on the counter." The second one gives three facts and a reason to come in now.
7. A process post from a voice note
Record 60 seconds on your phone while you shape loaves: what you're doing and why. Transcribe it (most phones and chat apps can) and paste the transcript in.
Here's a rough voice-note transcript of me explaining [process]:
[paste]
Turn it into a caption under 90 words for a photo of [what's in the photo].
Keep my phrases where they sound natural. Remove filler words.
Don't add facts that aren't in the transcript.
This is the prompt most likely to sound like you, because the words are yours. For a fuller method, see writing Instagram captions with AI that sound like you.
8. A seasonal pre-order announcement
We're taking pre-orders for [seasonal items, prices] for collection
[dates]. Order by [deadline] via [method]. Deposit: [amount or none].
Write: a caption (under 80 words), a counter-card version (under 40 words),
and a reminder caption for 3 days before the deadline.
Put the deadline in every version.
Check: dates and days of the week. Assistants sometimes pair a date with the wrong weekday. Compare against a calendar. A realistic slip: for festive pre-orders collected on 23 and 24 December 2026 with a deadline of Friday 18 December, the reminder caption read "Three days left: order by Monday 18 December". The 18th is a Friday, and a customer who believes the caption turns up on Monday to find the list closed. Give the assistant the date and the weekday together in the prompt, and ask it to repeat both exactly in every version, then check each against a calendar before posting.
Customer reply prompts
Replies are where AI saves the most time, because the same five or six messages arrive every week in slightly different words. Before pasting a customer's message into a chat assistant, remove anything you don't need to share: phone numbers, addresses, payment details. On consumer ChatGPT and Claude plans, you can also switch off the use of your chats for model training in privacy settings; business plans don't train on your content by default. Either way, the assistant only needs the question, not the customer's contact details.
9. An order enquiry that's missing details
Customer message: "[paste]"
Using my fact sheet, reply to ask only for the details I still need
to quote: [e.g. date, servings, flavour, collection or delivery, dietary needs].
Ask them as a short numbered list. Mention our notice period and deposit.
Under 100 words. Don't quote a price yet.
A customer writes: "Hi, do you do cakes for about 30 people? It's for my mum's 70th in a few weeks." A first draft (illustrative):
Hi [first name], we'd love to help make your mum's 70th
unforgettable! To put a quote together, could you tell me:
1. The date you need it
2. Your favourite flavours
3. Any dietary requirements
We need 10 days' notice for custom cakes.
Three fixes. "Unforgettable" and the exclamation mark belong on the fact sheet's never-say line. It dropped collection or delivery, which changes the price, so add it as question 4. And it forgot the deposit the prompt asked for; add "a 30% deposit secures the date". With the fact sheet's banned words in place, the next draft usually comes back without the gushing.
10. A complaint about a bake
Customer complaint: "[paste]"
What I know: [e.g. the batch was fine on our checks / we did overbake that tray].
What I'm offering: [replacement / refund / credit], per my complaints policy.
Draft a reply that: thanks them, says plainly what happened if we know,
makes the offer, and gives one way to reach me directly.
No "we're sorry for any inconvenience". Under 120 words.
Say a customer says a sourdough loaf was burnt on the base, and you know that tray did catch. Even with the banned phrase in the prompt, a first draft often opens "We're so sorry to hear about your experience. At our bakery, quality is our top priority." Both sentences could come from any business. The edited version keeps the facts: "Thank you for telling me. You're right: one tray on Tuesday caught on the base, and yours was from it. A fresh loaf is yours next time you're in, or I can refund the $6.50, whichever you'd prefer. I'm on the counter most mornings if you'd like a word." Specific, owns the mistake once, and makes the offer from your complaints policy rather than one the assistant invented.
If the complaint is a public review rather than a private message, the tone rules change: you're writing for everyone who reads it later, so keep the reply shorter, skip the details of the refund and invite the customer to get in touch directly.
11. An allergen question (the safe version)
Customer asks: "[paste]"
Our allergen policy: [paste from fact sheet].
The allergen information for [item] from our records is: [paste from your
allergen matrix, exactly].
Write a reply that states only what's in our records, repeats our
cross-contamination policy, and invites them to speak to staff before buying.
Do not add, guess or reassure beyond the records.
Notice what the AI is and isn't doing. It's arranging information you've supplied into a clear, polite reply. It isn't deciding whether the brownie contains nuts. If a customer asks about something not in your records, the answer is "let me check and come back to you", written by a person. The reasons are spelled out in whether an AI chatbot should answer allergen questions, and they apply to bakeries just as much.
12. A mix-up at collection
What happened: [e.g. customer arrived at 10am, order was set for 2pm;
or order had the wrong name piped on it].
Whose error: [ours / theirs / unclear].
What I can do: [e.g. ready in 40 minutes, redo the lettering, partial refund].
Draft a calm reply under 90 words. If it was our error, say so once, clearly.
If it wasn't, don't blame them; focus on the fix.
Checking what comes back
A 30-second check before anything goes out catches nearly every problem. Run through these five:
- Numbers. Every price, total, deposit and quantity matches your own sheet.
- Dates and times. The day of the week matches the date; hours match your current opening times.
- Promises. Nothing offered that you don't offer: no free delivery, no "gluten-free" or "nut-free" claims, no refunds beyond your policy.
- Invented detail. No ingredients, awards, origins or history you didn't supply.
- Voice. Read it aloud. If you wouldn't say it at the counter, change it.
When a prompt keeps giving you poor answers
Most bad outputs come from one of four causes, and each has a quick fix:
- Too long. Give a hard limit in words, not "short". "Under 40 words" works; "keep it brief" doesn't.
- Too gushing. Add banned words to the fact sheet and paste one of your own captions as an example. Say "match this tone".
- Wrong facts. The assistant filled a gap. Find the gap (a missing price, an unstated notice period) and add it to the fact sheet so it doesn't recur.
- Same structure every time. Ask for options in named styles, as prompt 5 does, or tell it which option you picked last week and ask for something different.
Don't start a fresh chat each time you correct something. Say what was wrong in the same conversation ("too many adjectives, and we don't open until 8am") and ask it to try again; then copy the fix into the fact sheet so the next conversation starts right.
How the prompts fit into a baker's week
Illustratively, say a two-person bakery with a small café counter uses them like this. On Monday afternoon the owner recosts one product using prompts 1 and 2, around 15 minutes including the spreadsheet check. On Wednesday, after the morning bake, she batches the week's captions with prompts 5 to 8 from three voice notes and six photos, about 25 minutes for five posts that used to take an evening. Customer replies happen as they arrive: paste, generate, check, send, usually under two minutes each instead of five to ten.
None of that is dramatic on its own. Across a month it frees a few hours for baking or rest, and the bigger gain is consistency: prices recalculated when costs move, posts that go out on schedule, and replies that give the same answer about deposits whoever writes them.
Saving the prompts so the whole team uses them
Copying prompts from a document works for one person. For a team, save the fact sheet and the prompts where everyone can reach them. ChatGPT and Claude both have Projects, and Gemini has Gems; each holds standing instructions and reference files so staff can type "reply to this cake enquiry" and get your format every time. Skip ChatGPT's custom GPTs: OpenAI is retiring them, and they stop running on 11 December 2026. Which of the three suits you depends on the assistant your team already uses; how projects and Gems compare sets out the differences.
Whichever you use, name one person to own the fact sheet. When butter prices or opening hours change, they update it the same day. A saved assistant working from last season's prices is worse than no assistant, because it sounds confident while being wrong.
Start with three prompts, not twelve: the costing prompt, one caption prompt and the enquiry reply. Use them for two weeks, tweak the wording where the answers annoy you, then add the rest.
Further reads
- How Bakeries Can Use AI to Predict Demand and Cut Unsold Stock — Use your sales data to decide how much of each bake to make.
- How to Build a Brand Voice Guide That AI Can Follow — Turn the fact sheet into a fuller voice guide AI can follow.
- How to Build a Shared Prompt Library for Your Team — Keep the team's best prompts in one place everyone uses.
- How to Write Menu Descriptions With AI That Still Sound Like You — Describe your bakes in your own voice, not stock food words.
- Social Media Pre-Publish Checklist for AI-Drafted Posts — A quick check before any AI-drafted post goes live.
- How to Draft Customer Email Replies With AI That Sound Like You — Longer email replies that still sound like you.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: ChatGPT and Claude plan and privacy pages (model-training settings on consumer and business plans), checked September 2026. Prices and costs in the examples are illustrative.