Running a cannabis delivery operation on Maui means a lot of writing. We update menus when inventory shifts, answer customers who want to know about delivery windows between Kahului, Kihei, and Lahaina, and write social posts that have to stay inside the compliance rules that apply to our license. Last year our team started testing AI writing tools to handle the repetitive parts. We also learned quickly that the gap between a useful draft and a generic one comes down almost entirely to the prompt. Along the way we even browsed collections of ai art prompts to see how other creators structure their instructions, which gave us a useful starting point for our own experiments.
Why prompts matter more than the tool
Most people blame the software when the output is bland. In our experience, the tool was rarely the problem. A vague request like ‘write a product description for our indica gummies’ produces filler: words like premium, relaxing, and perfect for any occasion. Those phrases sound fine until you realize they say nothing a customer can use, and they can drift into claims that don’t belong on a cannabis listing at all.
A working prompt usually includes four things:
- A role: who the assistant should write as, such as a plainspoken local shop assistant rather than a marketing copywriter.
- The audience: returning customers, first-time buyers, or visitors on vacation who only have a few days on island.
- Constraints: word limits, banned phrases, and the regulatory boundaries you must respect.
- An example of the finished tone: one short paragraph you wrote yourself, so the model has something concrete to match.
What we actually use
Our most-used prompts are dull on purpose. Dull prompts are repeatable, and repeatability is what a small operation needs. Here are the categories that earned a permanent place in our workflow.
Delivery window FAQs
Customers ask the same five questions every week: when orders cut off, whether we deliver to Upcountry, what ID is required at the door, and how to change an address. We wrote one prompt that takes our internal policy notes and produces a plain-language FAQ. Every time a policy changes, we paste in the new notes and regenerate the answers. Then a staff member reads each one before it goes live. The AI handles the phrasing; a human handles accuracy.
Menu descriptions
Product descriptions are where careless prompting does the most damage. Our rule is that the prompt must specify what the description may contain (strain type as listed by the producer, package size, flavor notes from the grower’s own sheet) and what it must never contain (health or medical benefit claims, dosage promises, or comparisons to prescription drugs). We include that restriction in every single prompt, not just once in a master document, because models forget instructions that sit far from the request.
Social posts and seasonal updates
For posts about holiday hours or new arrivals, we give the assistant a short list of facts and ask for three options at different lengths. Choosing among drafts is faster than editing one long draft. We also ask it to flag any sentence that could be read as an endorsement, and we review those flags ourselves.
How to judge whether a prompt works
A prompt that works is one you can hand to a coworker and get consistent results from. We test every new prompt against three inputs: a typical request, an edge case (a product with missing information, for example), and a deliberately tricky request that tempts the model to break a rule. If the prompt handles all three without babysitting, it goes into the library. If it needs heavy editing every time, we rewrite the constraints before we use it again. To go deeper, explore The marketplace for AI prompts that actually work.
We also keep a simple log: the date, the prompt version, what went wrong, and how we fixed it. After a few months, that log became more valuable than any single prompt, because it showed us which kinds of instructions our assistant reliably followed and which it ignored.
Building a prompt library for a small team
If you run a small business and want to start the same process, here is the approach that worked for us:
- Pick one repetitive task that costs you time every week.
- Write the prompt yourself first, with a role, audience, constraints, and a sample of the tone you want.
- Run it against real inputs from the past month, not invented examples.
- Save the version that works and label it with its purpose and date.
- Assign one person to review outputs before anything reaches a customer.
Resist the urge to build a library of fifty prompts in the first week. We started with three, retired two, and improved the third over several rounds. Depth beats volume.
Compliance is not optional
Cannabis rules differ by jurisdiction, and they change. Whatever tools you use, the responsibility for accurate, lawful content stays with your business. Treat AI output the way you would treat copy from a new hire: useful, fast, and in need of review. Keep a current copy of your license conditions and your state and local advertising restrictions next to your prompt library, and check drafts against them before publishing. If a prompt ever produces a claim you cannot support with your own documentation, delete the claim, not just the sentence around it.
What we would tell another island business
Start with the writing you dread, not the writing you enjoy. Write prompts the way you would brief a new employee, with specifics and clear limits. Keep a log of what fails. And remember that a good prompt is an operational asset, the same way a good delivery route or a reliable cooler is. It deserves maintenance.
Maui businesses already know how much depends on small details: a route that takes forty minutes instead of twenty, a cooler that holds temperature through a hot afternoon, a customer who needs a clear answer before they drive over the mountain. Prompts are the same kind of detail. Get them right once, test them honestly, and they quietly save hours every week.
Leave a Reply