Every few weeks our team gets the same question from a shop owner: can AI actually help a small cannabis delivery business, or is it just hype? The honest answer is that it can save real hours, but only when the prompts behind it are specific, tested, and written with your rules in mind. Many owners who try generative tools for the first time start with a vague request, get generic output, and give up. Others discover that a well-built prompt, whether for copy or for a visual concept, is a small asset that can be reused across the whole business. If you have experimented with image generators, you may have already searched for midjourney prompts and found a mix of gems and dead ends. This article is about separating the two.
Why most prompts fail for local businesses
A prompt that works for a travel poster or a fantasy illustration often falls apart when applied to a regulated product category. The main problems we see are predictable. The prompt is too short, so the model fills in gaps with clichés. It ignores audience and platform, so the copy reads like it was written for a national brand. It leaves out constraints, so the output includes imagery or phrasing that ad platforms and state regulators would never approve. And it is never tested, so nobody knows whether a slight change in wording makes a big difference.
For a delivery service, the stakes are practical. A product description that promises a specific effect is a compliance problem. A social image that looks like it is aimed at younger people is a problem. A text message with a misspelled product name is a customer service problem. A good prompt addresses these up front rather than hoping the editor catches them later.
What a usable prompt contains
After testing dozens of versions for our own use, we settled on a simple structure. Each prompt should state the role the model is playing, the audience, the channel, the length or format, the facts it must use, the things it must avoid, and an example of tone. That last part matters more than most people expect. Pasting in two or three sentences from your existing website or menu gives the model a voice to match.
Here is a stripped-down example for a product description. Rather than asking for a description of a pre-roll, you might write: act as a copywriter for a licensed delivery service, write a 60-word menu description using only the strain name, product type, and lab-tested potency values I provide, avoid any health, medical, or effect claims, avoid references to sleep, relaxation, or treatment, use a calm and informative tone, and end with a line reminding adults to check local rules. This is less exciting to write than a one-line request, but the output is far more consistent and far easier to approve.
Building a compliance-first creative workflow
Treat your prompts as part of your compliance process, not as a shortcut around it. Before any AI-generated asset goes live, run it through a short review. Check that the content does not depict or appeal to anyone under the legal age, does not use cartoon characters or candy-like styling, does not make health or therapeutic claims, and does not show consumption in public or while driving. Check that the packaging and product names match what your licensed products actually are. Check your state and local rules, because they differ and change over time, and when in doubt ask your regulator or counsel.
For imagery, we have found it more productive to ask for abstract or environmental scenes than for people or products. A prompt that requests a sunlit Kailua morning, ocean texture, a delivery van silhouette on a quiet street, and a clean typographic layout tends to produce something a platform reviewer is far less likely to flag. The goal is to convey place and service rather than lifestyle excess.
Prompts for the messages customers actually read
Some of the highest-value uses are unglamorous. Order confirmation texts, delivery window reminders, and reschedule notices all need to be clear and brief. A reliable prompt for these asks the model to produce three variants of each message, each under 160 characters, with the order number and window placeholders preserved exactly, and no marketing language inside transactional texts. Keeping transactional and promotional messages separate also helps with consent rules, so tell the model which category each message belongs to. To go deeper, explore The marketplace for AI prompts that actually work.
Email newsletters benefit from a different prompt. Give the model last month’s actual inventory categories, two upcoming hours or holiday schedule changes, and a single call to action. Ask for a subject line under 45 characters and a preview text that does not repeat the subject. Then edit heavily. The model is a drafting partner, not an author who knows your regulars.
How to test a prompt before you trust it
Testing does not need to be elaborate. Run each prompt three times with the same inputs and compare the outputs. If the results vary wildly in tone, the prompt is underspecified. If they all share the same clichés, add a list of banned phrases. Keep a simple log with the date, the prompt version, the model used, and whether the output needed major edits. Over a month, this log will show you which prompts are worth keeping and which need rewriting.
When you are comparing options, look for prompts that document their inputs, their intended use, and their known limits. A prompt that claims to work for everything is usually a prompt that works for nothing in particular. Specificity is the sign of something that has been used on real work.
A short checklist before you publish
- Does the prompt name the audience and the channel?
- Does it list the facts the output must use and forbid invented ones?
- Does it exclude health, effect, and treatment claims?
- Does the output avoid appeal to minors and any cartoon or candy styling?
- Have you run it at least three times with the same inputs?
- Has a person who knows your state rules reviewed the final copy or image?
- Is the prompt saved with a version number and a note about what it is for?
Where this leaves a local delivery business
AI tools will not replace a knowledgeable budtender, a careful compliance officer, or a driver who knows the neighborhood. What they can do is handle the repetitive drafting that eats time in a small shop, so your team can spend more energy on customer care and on the details that regulators check. The businesses that get value from these tools are the ones that treat prompts like standard operating procedures: written down, reviewed, improved, and retired when they stop working.
Start small. Pick one recurring task, such as weekly texts or menu descriptions, and build one strong prompt for it. Test it, log it, and refine it. Once that works, move to the next task. The payoff is not a flashy transformation but a steady reduction in the small frictions that slow a delivery operation down, and that is a result you can actually measure.
Leave a Reply