Every small delivery business eventually runs into the same wall: you need a lot of clear, friendly writing, and you need it fast. Product descriptions, delivery window texts, FAQ answers, review replies, and social posts all pile up, and most of us on the Maui team are not professional copywriters. That is why many of us started experimenting with AI tools, and why we began looking at an ai prompt marketplace as a place to find starting points instead of writing every instruction from scratch. This article covers what we learned, what failed, and how we now decide whether a prompt is worth keeping.
Why prompts matter more than the tool
When our first attempts came back, they were generic. We asked for a product description and got three paragraphs of vague enthusiasm that could have described anything from a coffee blend to a hiking boot. The tool was not the problem. The instructions were. A prompt that works is one that tells the model who the reader is, what the business does, what it must avoid, and what the output should look like.
Once we started treating prompts like small specifications rather than casual questions, the results improved quickly. Specificity was the single biggest difference.
What makes a prompt actually work
After several months of trial and error, our team settled on a short checklist that every prompt has to pass before it goes into our shared document:
- A defined audience. “Write for adults who already know the difference between indica and sativa” produces different copy than “write for a first-time customer.” We state which one we mean.
- Hard boundaries. We list what the output must not contain, such as health claims, medical promises, or language that could appeal to anyone under 21.
- A format. Word limits, bullet counts, tone, and whether to include a call to action. Vague format requests produce vague text.
- One real example. Pasting in a sample of copy we like gives the model a target to match far better than adjectives alone.
- A review step. No AI output goes live without a person reading it against our current product information and local rules.
Where we use prompts on the delivery side
Our day-to-day use falls into a few buckets. None of them replace human judgment, but they save real time.
Delivery window messages
Customers hate vague texts. We built a prompt that takes a driver name, an estimated arrival window, and the neighborhood, then returns a short message that is warm, plain, and specific. We removed any language that sounds like a promise of exact timing, because traffic on Maui can change a route in minutes, and we would rather under-promise.
First-time customer FAQ
New customers ask the same questions: how ID verification works, what happens if nobody answers the door, how to change an address, and whether they can pay in advance. We wrote the answers ourselves first, then used a prompt to adapt them into a consistent tone for our website and text templates. The rule was simple: the facts come from us, and the model only adjusts the wording.
Review responses
Replying to reviews is tedious, and generic replies look worse than no reply at all. Our prompt asks for a response that thanks the customer, addresses one specific detail they mentioned, and stays under four sentences. For negative reviews, it routes to a human and does not draft anything automatically.
Social and newsletter drafts
We use prompts to produce first drafts of seasonal newsletters, but we are strict about what is allowed in them. Anything that describes effects or suggests a use case gets cut. We have learned that the model will sometimes drift toward promotional language, so the prompt has to name those drifts directly. To go deeper, explore The marketplace for AI prompts that actually work.
Compliance is not optional
Cannabis marketing sits under rules that vary by jurisdiction, and those rules are not something a language model can reliably know or apply. Before any AI-generated text reaches a customer, a team member checks it against the current requirements from the Hawaii Department of Health and any applicable county guidance, and against our own licensing terms. If you run a similar business, treat this step as mandatory and document who signed off on what. Do not rely on a prompt to keep you compliant; a prompt is only as good as the person reviewing its output.
Building a small prompt library
The real payoff came when we stopped keeping prompts in private notes and started storing them in one shared place with names, versions, and a short description of what each one is for. A library lets a new staff member produce consistent messages on their first day. It also makes it obvious when a prompt has gone stale, such as after a menu change or a rule update.
Our library currently has around a dozen prompts, grouped by customer messages, product copy, internal notes, and marketing drafts. Each entry includes the date we last tested it and the name of the person who approved it. We review the whole set every quarter.
A simple test before you trust a prompt
Before adopting any prompt, run it three times with three different inputs. Read all three outputs. If they all sound like the same generic voice, the prompt needs more constraints. If one output breaks a rule you wrote down, add that rule explicitly and test again. A prompt that survives varied inputs is a prompt you can rely on; one that only looked good on the first try is not ready.
What we would do differently
If we were starting over, we would write our compliance boundaries and brand voice before writing a single prompt. We spent weeks refining prompts that had to be rewritten once we clarified what the business would and would not say. We would also involve the drivers earlier. They know which customer questions come up at the door, and their input made our FAQ far more useful than anything we wrote at a desk.
The bottom line
AI prompts are not magic, and a marketplace of them is only useful if the prompts are specific, tested, and reviewed. For a small cannabis delivery business on Maui, the value is not in replacing our voice but in keeping it consistent when we are busy, short-staffed, or answering the same question for the fifteenth time that week. Start with one workflow, write clear boundaries, test with varied inputs, and keep a person in the loop for anything a customer will read. That approach has served us better than chasing the cleverest prompt we could find.
Leave a Reply