When our team first started using AI image and copy tools for seasonal menu graphics, the early results were disappointing. The pictures looked generic, the captions sounded like every other storefront, and a few drafts broke our own compliance rules. What changed things was learning to write and reuse better instructions, and that is where a good library of midjourney prompts made a noticeable difference for us. The tool was never the bottleneck. The wording was.
What makes a prompt actually work
A prompt that works is one that produces a usable result on the first or second try, in a consistent style, without needing a long rescue session. For a delivery business, that usually means the prompt specifies four things clearly: the subject, the format, the audience, and the boundaries.
- Subject: what the image or text is actually about, such as a product category, a delivery window, or a holiday promotion.
- Format: the dimensions, the word count, the platform, or whether you want a list, a short email, or a caption.
- Audience: adult customers who are already verified, or staff who need an internal checklist.
- Boundaries: what the output must avoid, including health claims, depictions of consumption, and anything that could appeal to minors.
Vague prompts like “make a cool cannabis ad” produce vague, risky results. Specific prompts tend to produce work that needs only light editing.
Compliance comes first, not last
Cannabis marketing sits under strict rules that differ by state and sometimes by city. Before any AI-generated text or image goes live, it should pass a human review against your state’s advertising regulations and your own policies. We build that review into the prompt itself wherever we can. For example, every copy prompt we use now includes a line stating that the output must not make medical or therapeutic claims, must not promise specific effects, and must not include age-appealing imagery or characters.
Building the rules into the prompt does not replace review. It reduces the number of drafts that need heavy correction, and it gives new staff a clear standard to follow. Keep a short written policy next to your prompt library so anyone can see why each guardrail exists.
The prompts we use most
Our day-to-day work falls into a handful of categories. Each has a template we refine over time.
Product description drafts
Descriptions for flower, pre-rolls, edibles, and accessories need to be accurate to the lab results and packaging. Our template asks the model to work only from the facts we paste in, such as strain type, THC and CBD percentages from the certificate of analysis, package size, and serving information. The model is told to write in plain language and to avoid comparisons with other businesses. We then check every number against the source document by hand.
Delivery-window and order-status emails
Customers respond well to short, clear messages about when their order will arrive. A good prompt specifies the maximum length, the tone, the required fields, and the legal footer. We keep separate versions for confirmation, out-for-delivery, and delayed orders, so the tone stays calm even when a driver is running late.
Social captions and seasonal promotions
Social platforms have their own restrictions on cannabis content, and those rules change. Our caption prompts ask for educational or operational messages only, such as delivery hours, holiday schedules, or new local partnerships, and they explicitly exclude product imagery that might not pass platform review. We also ask for alternate versions so we can choose the most conservative option before posting.
Internal training and SOPs
Some of the most useful prompts are internal. We use them to turn a messy list of verification steps into a checklist, or to draft a short scenario that helps new drivers practice refusing a sale to someone without valid ID. Those prompts rarely go public, but they save time and reduce errors.
How to evaluate a prompt before you rely on it
Whether you write prompts yourself or buy them from a marketplace, test them the same way. A prompt earns a place in your library only after it has passed a few simple checks.
- Run it three times with the same input. If the outputs vary wildly in tone or accuracy, the prompt is underspecified.
- Run it with an edge case, such as a product with unusual dosing information or a delivery delayed by weather. Watch whether the model invents details.
- Check every factual claim against your source material. Never let a model supply lab numbers, legal language, or licensing details from memory.
- Have someone outside the marketing role read the output. A driver or customer service rep often catches wording that sounds fine to the person who wrote the prompt.
- Record the version number and the date. Rules change, and you need to know which prompt produced which piece of content.
Where to find prompts worth paying for
Free prompts are plentiful, but they are often untested and written for generic use cases. If you decide to buy prompts rather than write them all in-house, look for listings that describe the input required, show sample outputs, and state clearly what the prompt was designed to do. A marketplace such as PromptMart’s collection of tested AI prompts can be a useful place to compare options, as long as you still run your own compliance review on anything you adapt for a regulated business.
Price should not be your only filter. A cheaper prompt that needs ten minutes of rewriting every time costs more in staff hours than a well-documented one that works out of the box.
Building a prompt library that survives staff turnover
The biggest risk with AI prompts in a small business is that they live in one person’s notes app. When that person leaves, the institutional knowledge goes with them. We keep ours in a shared document organized by category, with each entry listing its purpose, required inputs, guardrails, last review date, and the name of the person responsible for updating it.
We also keep a short “retired” section. When a prompt produces a result that later turned out to be inaccurate or non-compliant, we move it there with a note explaining why. That history has been more valuable than we expected, because it shows patterns in what the models get wrong.
Keeping humans in the loop
AI tools are fast, but they do not understand your state’s rules, your licensing conditions, or the reality of a delivery route at 9 p.m. on a holiday weekend. Every output that reaches a customer, a platform, or a regulator should have a named human who approved it. Build that approval step into your workflow from the start rather than adding it after a problem appears.
For delivery businesses in particular, the goal is not to automate customer trust. It is to free up time so your team can spend more of it on the things customers actually notice: accurate orders, respectful drivers, and clear communication when something changes.
A simple starting plan
If you are just beginning, do not try to build everything at once. Pick two repetitive tasks, such as order-status emails and weekly social posts. Write or buy one prompt for each, add your compliance guardrails, test them using the checklist above, and log the results for a month. Once those two are reliable, expand into product descriptions and internal training. Small, documented steps tend to produce a library your whole team can trust.
Leave a Reply