Every week, our dispatch and customer service team in Winnipeg answers the same handful of questions: how delivery windows work, what happens if nobody is home at the door, and whether an order can be changed after it has been placed. Lately we have started using AI writing tools to draft answers faster, and the biggest difference has come from the prompts themselves, not the tool. If you are looking for best ai prompts that have been tested and organized instead of guessed at, a marketplace built around that idea is a sensible place to begin.
Why prompt quality matters more than the tool
A vague request like ‘write a reply about delivery times’ produces a vague answer. The model has no idea whether you are a licensed retailer, what your cutoff hours are, or what tone your customers expect. A useful prompt supplies that context up front, so the output needs far less editing.
In practice, the prompts that work for us share a few traits:
- They name the role the assistant plays, such as customer service agent for a licensed delivery service.
- They include the actual policy text the answer must follow.
- They forbid specific things, such as guessing at delivery times or making health claims.
- They specify the output format, whether that is a short SMS, a two-line email or a bulleted checklist.
Where AI genuinely helps a cannabis delivery business
Not every task deserves an AI assistant. The jobs where we have seen real time savings are repetitive and text-heavy, and where a person still reviews the final message before it goes out.
- Customer FAQs: drafting answers about delivery zones, ID checks at the door, payment options and substitutions.
- Order status messages: turning a dispatch update into a friendly, plain-language text.
- Staff onboarding: generating practice scenarios, such as a customer who is underage or an address outside the service area, so new drivers can rehearse the right response.
- Review responses: producing a first draft of a reply to a public review, which a manager then edits.
- Internal summaries: condensing long policy documents into a one-page checklist for the shift lead.
The compliance line you cannot prompt your way around
This is the part that matters most for anyone in this industry. Under Canada’s Cannabis Act, promotion of cannabis is tightly restricted. In general, it cannot be appealing to young people, cannot describe a lifestyle or glamorize use, and cannot make false or misleading statements, including about health or effects. Manitoba also has its own rules for retail and delivery operations. Before you publish any AI-drafted copy, check it against current federal and provincial requirements, and if you are unsure, confirm with a licensed legal advisor.
The practical lesson is to build those rules into the prompt itself. For example, we tell the assistant that it must not describe effects, must not use slang aimed at new users, and must suggest that the customer consult their pharmacist or a health professional if they ask about medical use. The model will not catch every problem, so a human review step is still required.
Building a prompt library for your team
One-off prompts get lost quickly. Once a few people on the team are writing their own versions, your replies start to sound inconsistent and nobody remembers which version is approved. We now keep a shared document with a small number of approved prompts, each labeled with its purpose and the last date someone reviewed it.
Step 1: Write the role and the boundaries
Start with a single paragraph that describes who the assistant is and what it must never do. Keep it specific to your operation rather than generic.
Step 2: Paste in the policy, not a summary
If your delivery window policy says orders placed after a certain hour roll to the next day, paste that exact wording into the prompt. Summaries drift, and the model will fill gaps with confident-sounding guesses.
Step 3: Set the output format
Tell the assistant how long the reply should be, what tone to use and whether to include a sign-off. Consistent formatting makes it easier for staff to review quickly.
Step 4: Test with awkward cases
Try the prompt on the questions that cause the most trouble: a refund request after a missed delivery, a customer asking whether a product is safe to combine with medication, a message written in French. Note where the output fails and tighten the prompt until it behaves.
When you are comparing collections, look for prompts that document what they were tested against and where they fall short. A curated library of tested AI prompts usually explains those trade-offs more clearly than starting from a blank chat window, and that clarity saves time when you are adapting a template to your own rules.
A sample prompt for our FAQ work looks something like this: ‘You are a customer service assistant for a licensed cannabis delivery service in Winnipeg. Answer only using the delivery policy provided below. Do not describe effects, do not make health claims, and do not guess at delivery times. If the answer is not in the policy, tell the customer that a team member will follow up. Keep replies under 80 words and end with our support hours.’
Keeping humans in the loop
AI drafts should never go straight to a customer, especially in a regulated industry. Our rule is simple: the assistant writes, a trained team member reads, and only then does the message go out. Anything involving age verification, medical questions, complaints or legal topics is escalated to a manager rather than answered from a template.
This also protects the quality of the work. People who edit AI drafts regularly start to notice patterns, such as a prompt that always sounds too formal or one that invents a detail, and they feed that back into the shared library.
A simple first-week plan
- Day one: list the five questions your team answers most often.
- Day two: write one prompt for each question, pasting in the exact policy text.
- Day three: test each prompt against two awkward cases and record the failures.
- Day four: have a manager review every compliance-sensitive prompt against current regulations.
- Day five: publish the approved prompts in a shared folder with a review date on each one.
Used this way, AI becomes a drafting assistant that makes your team faster without making your messaging riskier. The goal is not to automate customer care but to give your people a better starting point, so they can spend their time on the conversations that actually need a human.
Leave a Reply