Every week our team faced the same chore: refresh the delivery menu, write notes for new arrivals, draft a text to loyal customers about weekend specials, and answer the same questions about order windows. Each task was small, but together they ate hours we needed for packing orders and checking on drivers. When we started looking for a better way, we found that the difference between a useful AI answer and a useless one often came down to the instructions we gave it. That led us to explore prompt templates for marketers, which are reusable instructions that spell out the role, the facts, the tone and the limits before the AI writes a single word.
What makes a prompt actually work
Most of our early attempts failed for the same reasons. We typed things like ‘write a menu description for a sativa flower’ and got generic copy that could have described any product in any shop. A prompt that works tends to include a handful of specific elements:
- A clear role: for example, ‘You are a copywriter for a local delivery service that serves adult customers.’
- Verified facts only: strain name, growing method, lab-tested potency and terpene notes copied from the product sheet, with an instruction not to invent any of them.
- A defined output format: a 40-word description, three bullet points, or a 160-character text message.
- Tone boundaries: friendly and plain-spoken, never slangy in a way that could appeal to minors.
- Banned claims: an explicit list of phrases the AI must avoid, such as health benefits or medical outcomes.
Once we added these pieces, the drafts became much more usable. They still needed editing, but they stopped sounding like a template pulled from the internet and started sounding like us.
Where a prompt library fits in
We did not want to build every prompt from nothing, and we did not want to copy random examples from forums. Instead, we looked at curated collections where prompts are written for specific marketing jobs and then reviewed for clarity. A marketplace such as a curated library of tested business prompts made it easier to compare structures and borrow the parts that fit our shop. The value was not the prompt itself so much as the pattern: a clear role, a fixed output length and a checklist of things to verify before publishing.
Our rule is simple. A downloaded or purchased prompt is a starting draft, not a finished workflow. We rewrite the placeholders with our own product data, test the output on three or four real items, and keep a note of which versions produced copy our staff would actually sign off on.
Compliance comes before creativity
Cannabis marketing carries constraints that most industries do not face, and no prompt can replace knowing your local obligations. Before we let any AI tool touch customer-facing copy, we wrote down the rules we follow and built them into every prompt that touches marketing.
Our standing rules for AI-drafted copy
- No claims about curing, treating or preventing any condition.
- No imagery or language aimed at people under the legal age, including cartoon styles or youth slang.
- Age-gate language and license information included on every page that describes products.
- Texts and emails only sent to customers who have opted in, with a clear way to unsubscribe.
- Every AI draft is read by a human who checks the facts against the product label before it goes live.
We also check the current rules in our state and municipality regularly, because advertising restrictions can change. If a rule is unclear, we ask our licensing contact or a qualified attorney rather than guessing. The AI does not know your jurisdiction, so the prompt should tell it what to avoid and the reviewer should confirm the result.
Our menu description workflow
The menu was our first real test. Here is how the process runs each week:
- The inventory lead exports the new products with strain, potency, weight and supplier notes.
- A staff member pastes that data into the prompt template, along with the banned-claims list and the required word count.
- The AI returns three short descriptions per product.
- A second staff member compares each draft to the label and removes anything that goes beyond the facts.
- The approved copy is loaded into the menu, and the old version is archived so we can compare performance later.
This process takes about the same time as writing from scratch for a single product, but it scales much better. On a week with fifteen new items, the saving is obvious. We also noticed that consistency improved. Our descriptions now share a structure, so customers can scan them quickly and compare options.
Customer texts and the opt-in question
Text messaging is where we were most cautious. A well-written message can bring back a customer who has not ordered in a month, but a careless one can cause complaints or worse. Our prompt for texts includes three instructions that we treat as non-negotiable: the message must be under a set character limit, it must include an unsubscribe option, and it must never reference a customer’s personal habits or purchase history in a way they would not expect.
We also ask the AI for two versions of each message, one informational and one promotional, so that we can choose the calmer option when something sensitive is happening, such as a holiday closure or a change to delivery hours. The informational version is often the one we send, because customers tend to trust a message that simply tells them when we are open.
What we still write ourselves
AI has not replaced the parts of our business that depend on local knowledge. We still write the welcome message for new neighbors in our delivery zone, the apology when a route runs late because of weather, and the notes that explain our driver safety practices. Those messages carry our voice and our accountability, and a template cannot know that a particular street has had road work for three weeks or that a regular customer prefers a call before arrival.
We also keep a short list of phrases that feel too polished. When an AI draft uses language that sounds like a press release, we edit it into plainer speech. Our customers are neighbors, and they respond better to direct sentences than to marketing gloss.
A starting checklist for small delivery businesses
- Write down your compliance rules before you write any prompt.
- Build one template per task: menus, product notes, texts, and FAQ answers.
- Put verified product data into the prompt, and forbid the AI from adding facts that are not there.
- Test each template on real items and keep the versions that work.
- Require a human review against the label or policy before anything is published.
- Review your templates every few months, especially after rule changes or menu shifts.
The honest takeaway
AI prompts are not magic, and they are not a shortcut around the responsibilities that come with selling a regulated product. What they have given us is a faster first draft and a more consistent structure, which leaves our staff more time for the work that matters: careful packing, safe deliveries and answering customer questions with real knowledge. If you run a small cannabis delivery service and you are considering this approach, start with one task, write your rules first, and treat every output as a draft that a person must approve.

Leave a Reply