Anyone running a cannabis delivery operation in Quebec knows the daily grind: order confirmations in French and English, product questions that need careful wording, driver scheduling, and customer messages that must stay within advertising limits. When our team started testing AI tools to speed up this work, we quickly learned that the search for the best ai prompts is less about finding clever phrasing and more about finding instructions that respect your industry, your language rules, and your customers. This article explains what we learned, what we kept, and what we threw out. If best ai prompts is what brought you here, start with the guide below.
Why generic prompts fail in a regulated delivery business
Most prompt lists online are written for marketing copywriters or software developers. They ask for punchy social posts or code snippets. A cannabis delivery team needs something different: a model that writes neutral, factual product descriptions, avoids health claims, and never implies that a product treats a condition. A prompt that asks for a “energetic, relaxing strain description with benefits” will produce exactly the copy you should not publish.
The first lesson we took from our testing was simple. Every prompt we use now includes three things: the role the assistant plays, the legal or policy boundaries it must follow, and the format we expect back. Without those three elements, the output looked polished but needed heavy editing, and sometimes needed to be discarded entirely.
Our three-part prompt structure
After several rounds of revisions, our internal templates follow the same pattern:
- Role: Tell the assistant it is writing for a licensed-style delivery service in Quebec, serving adults only, with no medical or therapeutic language.
- Boundaries: List the words and claims to avoid, such as cure, treat, guaranteed effect, or any suggestion of targeting minors. Ask it to flag anything it is unsure about rather than guessing.
- Output format: Request a fixed structure, for example a two-sentence summary, a bulleted list of product facts supplied by us, and a separate French version written naturally rather than translated word for word.
We also ask the model to say when it does not have enough information. This single instruction has saved us from publishing invented terpene percentages or made-up lab results, which is exactly the kind of error a regulator or a frustrated customer would notice.
Prompts for customer communication
Order updates are the highest-volume task in our business, so they were our first target. A useful prompt for these messages gives the assistant the order status, the estimated window in plain terms, and the tone we want: calm, short, and respectful. We instruct it to never mention the contents of a order in a way that could be seen by someone else on a shared phone or email, and to keep identifying details out of notifications.
For bilingual replies, we found it better to ask for the French version first, then request an English version that keeps the same meaning. Asking for both at once often produced stiff sentences in one language. Quebec customers notice when French reads like a translation, so this order matters.
Prompts for product catalogue work
Product descriptions are where careless prompting creates real risk. Our approach is to supply the verified facts ourselves, including the product name, format, weight, and producer information, then ask the assistant only to rephrase those facts clearly. We never ask it to invent flavour notes or effects. If a field is missing, the prompt tells the model to leave a placeholder for our staff to complete.
This keeps the assistant in the role of editor rather than source. It also makes review faster, because our staff check a short list of facts instead of reading a full essay for hidden claims.
Prompts for driver and dispatch support
Dispatch is less glamorous but equally valuable. We use prompts that turn messy notes from drivers into clean incident summaries, and that draft polite delay notices when traffic on the Pont Champlain or a winter storm slows a route. The key constraint here is privacy: the prompt instructs the assistant to remove addresses, phone numbers, and names from anything that will be shared beyond the dispatch team.
How to judge whether a prompt actually works
Our team uses a simple checklist before any prompt goes into regular rotation:
- Does the output need fewer than two edits to be publishable?
- Does it respect our language and policy boundaries on every test run, not just the first?
- Would a new staff member be able to reuse it without extra explanation?
- Does it say when information is missing instead of filling gaps?
A prompt that fails any of these questions goes back to the drawing board. We keep a shared document with the versions we retired and the reason for each change, which has turned out to be one of the most useful internal resources we have built.
Where to find a starting point
Building everything from scratch took our team many weeks. If you are at the beginning of this process, it can help to review tested examples before writing your own. One option we have looked at is a curated prompt marketplace where submitted prompts are organised by use case, which makes it easier to compare structures side by side. Treat any downloaded prompt as a draft: adapt the boundaries to your jurisdiction, remove any wording that drifts toward health or performance claims, and test it with your own sample data before it reaches a customer.
A note on compliance and common sense
AI tools are not a substitute for legal advice. Cannabis advertising and sales rules in Canada and Quebec are detailed and change over time, so confirm current requirements with qualified counsel or the relevant provincial authority. Our prompts reflect the restrictions we have reviewed with our own advisers, and we still have a human approve anything customer-facing.
Getting started this week
If you want to put this into practice, start small. Pick one repetitive task, such as order status messages, and write a prompt with a role, boundaries, and a fixed output format. Run it ten times with realistic inputs, log every failure, and revise. Only after it passes consistently should you expand to catalogue work or dispatch. Teams that rush this stage tend to spend more time correcting output than they saved by using the tool.
The goal is not to hand your business to a machine. It is to give your staff clearer starting points so they spend less time on routine drafting and more time on the parts of delivery that depend on judgment, local knowledge, and care for the customer. Prompts that work are the ones that keep those priorities in view.
Leave a Reply