Every week our team writes the same things over and over: order confirmations, out-of-stock notices, a fresh description for a new product, a reminder that ID is checked at the door, and answers to the questions customers ask about delivery windows. When we first tried AI tools to speed this up, the drafts were a mess. They were too cheerful, sometimes too specific about effects, and occasionally written in a tone that would have been a problem for a regulated business. What changed our results was not a better tool. It was learning to write better prompts, and finding a library of prompt templates for marketers gave us a starting structure we could adapt to our own rules.
Why generic AI output fails in a regulated niche
Most prompts people share online are written for ordinary consumer brands. They ask for punchy copy, emotional hooks and a call to action. Those instructions are exactly what gets a cannabis business into trouble. Promotion of cannabis in Canada is governed by federal rules under the Cannabis Act and its regulations, and Quebec adds its own requirements on top. Language that appeals to young people, lifestyle imagery, and unverified claims about wellness or health are areas where a retailer can get into difficulty quickly.
A generic prompt also has no idea what your business actually sells. If you ask an AI model to describe a product without supplying the facts, it will invent them. Terms like “calming” or “energizing” can appear without any basis in the product information you hold, and that invented language is the most dangerous output you can publish.
Important: This article describes our workflow and is not legal advice. Cannabis advertising rules change, and any copy you publish should be reviewed by a qualified lawyer or compliance advisor familiar with current federal and Quebec requirements.
The four parts every compliant prompt needs
After many rounds of testing, we settled on a prompt structure with four parts. Each one solves a specific failure we saw in early drafts.
- Role and audience: Tell the model it is writing for an adult-only service and that the reader is already a verified adult customer looking for practical information, not an audience to be persuaded.
- Source facts only: Paste in the product data from your own catalogue, such as THC and CBD percentages, format, weight and batch notes. Then instruct the model to use only those facts and to write “information not provided” wherever a detail is missing.
- Prohibited content list: Spell out what must never appear. For us that includes health or therapeutic claims, references to sleep, anxiety or pain relief, any imagery or language aimed at minors, slang that appeals to youth, and comparisons to alcohol or tobacco.
- Output format: Specify length, headings, reading level and whether you want a plain factual tone. Constraining the format cut our editing time more than anything else.
Example: a product description prompt
Here is the general shape of the prompt we use for new product listings. We have removed the specifics of our catalogue, but the logic is what matters:
Act as a copy editor for a licensed adult-use delivery service in Quebec. Write a product description of no more than 90 words using only the facts in the data block below. Do not describe effects, health benefits or intended use. Do not use lifestyle scenes, humour aimed at younger people, or superlatives such as best or strongest. If a field is empty, state that the information is not provided. End with a reminder that customers must be of legal age to purchase.
The result is dull, and that is the point. Dull, factual copy is far easier to defend than energetic copy that has to be walked back later. Our compliance reviewer now approves most drafts on the first pass.
Building a library instead of rewriting from scratch
The real efficiency gain came when we stopped treating each prompt as a one-off. We now keep a shared document with separate versions for product listings, order status messages, out-of-stock notices, delivery-window FAQs and reply templates for common complaints. Each version has a note on when it was last reviewed and who approved it.
Consistency also matters for customers. When someone receives a delivery update, the wording should match what they read on the website and what the driver says at the door. A shared prompt library helps keep those messages aligned across channels.
Where a prompt marketplace helps
Some teams prefer to build everything in-house, and that is reasonable if you have someone who understands prompt design well. For smaller teams, a curated marketplace can be a faster place to begin. The useful question is not whether a prompt is popular but whether its author explains the inputs, the expected output and the cases where it fails. If you are exploring that option, a directory of tested prompts for specific business tasks can show you how other people document their instructions, which is a useful model even if you never buy anything.
What to avoid when adopting someone else’s prompt
- Prompts written for alcohol, supplements or general retail that were never adapted to cannabis rules.
- Prompts with no stated inputs, which usually means the model is filling gaps with guesses.
- Prompts that encourage a “friendly” or “hype” voice without any guardrails on claims.
- Prompts that generate social media captions without a review step for imagery and hashtags.
A review checklist before anything goes live
Before we publish any AI-assisted copy, we run it through the same short checklist. It takes a few minutes and has caught problems every month.
- Does every factual statement match a verified product record?
- Is there any health, wellness or therapeutic language, including implied language?
- Would the copy appeal to someone under 21 or use imagery and tone that does?
- Is the age-restriction notice present where required?
- Has a named person approved the final version, with the date recorded?
Keeping humans in the loop
AI speeds up drafting, but it does not replace judgment. Our driver feedback, complaint logs and customer questions are now the main input for updating prompts. When a customer asks something the FAQ did not answer, that question becomes a new prompt to write. Over time the library gets more specific to our customers and more accurate about what we can and cannot say.
The takeaway
AI tools can save a small cannabis delivery team a lot of time on routine writing, but only if the prompts carry the rules that a generic prompt leaves out. Give the model your facts, forbid the claims you cannot support, define the output, and keep a named human responsible for approval. Start with one or two high-volume tasks, such as order updates or product listings, refine the prompts through real use, and expand from there. Speed is useful, but in this industry the discipline behind the prompt matters more than the speed.
Leave a Reply