Most of us on the team first tried AI writing tools the same way: we typed a vague request, got a bland answer, and closed the tab. Anyone who has searched for ai art prompts online has seen how much the wording changes the result, and the same is true for text. The difference between a lazy request and a useful one is often just a few clear details. For a cannabis delivery business in Vancouver, where menus change weekly and every customer message has to be accurate and compliant, that difference adds up quickly.
Where a delivery shop actually needs help
Our day-to-day writing is less glamorous than marketing campaigns. We update menu listings when a new batch of edibles arrives, answer repeated questions about delivery windows, write order confirmation texts, and keep the FAQ page current when rules or hours change. None of these tasks is hard, but they eat time, and they are easy to get slightly wrong when you are rushing.
That is where a good prompt earns its place. A well-built prompt can turn a messy spreadsheet of product details into clean menu descriptions, or convert a long policy document into a short FAQ that a customer can read on their phone while waiting at the door.
What a prompt that works looks like
After a few months of trial and error, we noticed that the prompts that produced usable output shared the same structure. They were not clever. They were specific.
- A defined role: “You are writing for a licensed delivery service in Metro Vancouver. Your tone is friendly and plain-spoken.”
- Context: The product details, the audience, and what the text is for, such as a menu card or a text message.
- Hard limits: Words to avoid, claims that are not allowed, and the maximum length.
- An output format: A bulleted list, a 40-word blurb, or a table with fixed columns.
- A sample of the house style: One or two examples of text you already like, pasted in for reference.
Once we started including all five elements, the first draft was usually close enough to edit rather than rewrite. That alone saved our staff a surprising amount of time during busy evenings.
The compliance check comes first
This is the part that is specific to our industry, and it is not optional. Cannabis promotion in Canada is regulated, and in British Columbia the rules for licensed retailers and delivery services cover what you can say, where you can say it, and who you can appeal to. Restrictions generally include avoiding messaging that appeals to young people, avoiding unsupported health or medical claims, and being careful with anything that looks like an inducement. Rules can change, and your licence conditions may add requirements of their own.
We made a rule that no AI-generated text goes live until a person has checked it against the current provincial guidance and our own licence conditions. We also added a standing instruction to every prompt: do not make health claims, do not refer to effects in a way that could be read as medical, and do not use language aimed at younger audiences. The model does not always follow these instructions perfectly, so the human review step is what actually protects us.
Keeping your voice while using a prompt library
One concern we heard from other owners is that AI copy all sounds the same. That is a fair worry. A reasonable middle path is to start from a tested structure rather than a blank page. If you want a head start, a marketplace of reusable instructions such as the PromptMart collection of prompts for writing and design work can give you a framework to adapt to your own brand, which is much faster than inventing a structure from scratch every time.
The key word is adapt. We keep a shared document with the prompts we use most, each annotated with notes on what changed the output for better or worse. Over time that document has become more valuable than any single prompt, because it reflects how our customers actually talk and what they ask about.
A sample workflow for a weekly menu update
Here is the process we currently follow each week. It takes about forty minutes, down from the couple of hours it used to take.
- Export the new product list with names, weights, product types, and any lab-confirmed potency figures.
- Paste the list into the menu prompt, along with the house style example and the compliance instructions.
- Ask for short descriptions only, with no added claims beyond what the product sheet says.
- Read every description against the source sheet. Correct any detail the model guessed at.
- Have a second staff member review the final copy before it goes to the website and the in-app listings.
Never let the model invent the numbers
This deserves its own section because it is the most important mistake to avoid. Potency figures, strain classifications, and ingredient amounts must come from the lab certificate or the supplier sheet, never from the model. If a field is missing, the prompt should tell the model to leave it blank or to flag it, rather than fill it with something that sounds plausible. We now put that instruction in every prompt that touches product data, and we spot-check outputs against source documents every week.
Final thoughts
AI tools are useful for a delivery business in the same way a good template is useful. They save time on repetitive writing and help you start faster, but they do not replace knowing your products, your customers, or the rules that govern what you are allowed to say. Treat prompts as drafts, keep a human in the loop for anything customer-facing, and build a shared library of what works for your shop. Done that way, the time you save goes back into the things that matter most: accurate orders, fast deliveries, and a team that knows what it is talking about.
Leave a Reply