Most small delivery services don’t have a designer, a copywriter, or a spare afternoon to build a month of social posts. That gap is why many owners have started experimenting with AI image and text tools, and why people searching for midjourney prompts have become such a common sight in shop owner forums. The tools are only as useful as the instructions you give them, and a vague request usually produces vague, off-brand, or occasionally non-compliant output.
This article walks through how our team at a Newport-area cannabis delivery service thinks about prompts, what we have learned about separating useful ones from noise, and where the compliance lines sit for a business in our industry. Nothing here is legal advice, so check current Rhode Island rules with your licensing counsel before you publish anything.
Why a delivery service needs prompts at all
Our day is mostly logistics. Orders come in, drivers need routes, inventory changes daily, and the menu has to reflect what is actually on the shelf. Marketing gets whatever time is left over. A few recurring tasks eat that time quickly:
- Rewriting product descriptions when a new batch arrives from a supplier
- Drafting a weekly email to existing customers about hours, holiday schedules, or new categories
- Producing simple concept art for a seasonal banner or a loyalty card design
- Writing short scripts for staff who answer phones during peak evening hours
A good prompt can turn a 40-minute writing task into a five-minute edit. A bad prompt creates more cleanup than it saves. That difference is the whole reason prompt quality matters to us.
What “actually works” means in practice
When we first tried prompt libraries, we judged them by how impressive the sample outputs looked. That turned out to be a poor test. A striking image or a witty paragraph in a demo tells you almost nothing about whether the prompt will hold up when you swap in your own product names, your own tone, and your own constraints.
We now use a simpler standard. A prompt works if it meets four conditions:
- It produces a usable first draft without heavy rewriting
- It accepts variables cleanly, such as strain category, pack size, or neighborhood name, without breaking the output
- Its results stay consistent across several runs
- It states its constraints clearly, so the model knows what to avoid
That last point matters more than people expect. Constraints are where most of the value lives. A prompt that says “write a product description” will drift into health claims or lifestyle imagery. A prompt that says “write a factual, two-sentence description of the product’s listed ingredients and format, avoid medical language, and do not mention effects or use by minors as an appeal” gives you something you can actually publish after a read-through.
Compliance-first prompting
This is the part other delivery services and dispensaries should take seriously. Cannabis advertising is heavily restricted in most places, and the rules are specific about audience, claims, and imagery. Rhode Island’s adult-use framework has its own requirements, and enforcement details change, so treat everything below as a starting checklist rather than a settled rulebook.
Build restrictions into every prompt
We keep a short block of standing instructions that gets pasted into every prompt we use for customer-facing copy. It covers the basics: no claims about medical benefits, no language aimed at people under 21, no depictions of consumption, and no suggestion that a product is safer or stronger than another without verifiable lab data. Copying that block takes ten seconds and removes a whole category of mistakes.
Be careful with image prompts
Image generators are where compliance problems hide. A request for “relaxed people enjoying a sunny afternoon with their favorite flower” can easily produce imagery that looks like it targets a young audience or shows consumption. We ask for neutral product photography styles, packaging mockups, and abstract textures instead. If an image could be mistaken for a lifestyle ad aimed at a broad audience, we do not use it.
Keep a human review step
No prompt replaces a person reading the final output against your state’s rules and your own brand standards. We assign one staff member to review anything that goes on the website, in email, or on social media. That review takes minutes, and it is the reason we have never had to pull a post because of an AI-generated claim.
Using prompts for menu and product copy
Menu copy is the most useful place to start, because the inputs are structured. Every product has a name, a category, a weight, a lab-reported potency, and a supplier. A well-built prompt treats those fields as variables and asks for a fixed output format. To go deeper, explore The marketplace for AI prompts that actually work.
Here is how we structure a menu description request in plain terms:
- Provide the verified fields exactly as they appear on the lab report
- Specify the word count, usually one or two short sentences
- Name the banned phrases and topics explicitly
- Ask for a plain, informative tone that matches the rest of the menu
- Request three variations so you can pick the most accurate one
The three-variation step is underrated. Often the second or third option is clearer than the first, and comparing them quickly shows you where the model was guessing.
Using prompts for email and staff scripts
Customer emails work well with prompts when you give the model a specific reason to write. “Announce our holiday hours” is fine. “Announce that we will close at 9 p.m. on the 24th and reopen at 10 a.m. on the 26th, mention that orders placed before close will be fulfilled the same day, and keep the message under 120 words” is better. The second version leaves less room for invented details.
Staff scripts are another practical use. We generate short call-handling scripts for common questions: delivery windows, minimum order amounts, how to verify identity at the door. We then rewrite them in our own voice and run them past the team. The prompt gives us a skeleton, and the people who answer the phone supply the judgment.
Vetting a prompt marketplace before you buy
If you decide to buy prompts rather than write your own, look at the marketplace the way you would look at any software vendor. Some questions we asked before we spent a dollar:
- Are prompts described with example inputs and outputs that you can reproduce?
- Does the seller explain what the prompt is designed for, and what it is not designed for?
- Can you see how a prompt handles variables, or only a single polished example?
- Is there a way to report prompts that produce poor or unsafe output?
- Are the terms clear about licensing for commercial use of generated content?
That last question matters for any business that publishes content publicly. Confirm that you are allowed to use what you generate for commercial purposes, and keep records of the prompts you used in case a platform or partner asks later.
Where we stopped using AI prompts
We have also learned what not to delegate. We do not use generated text for anything that describes effects, dosing, or health outcomes. We do not use it to answer customer questions about medical use. We do not let it write our privacy notices or our terms of service, because those need to match our actual data practices and our licensing obligations. In each case, a prompt can help draft a starting structure, but a qualified person has to own the final version.
A simple workflow to start with
If you run a small delivery service and want to try this without making a mess, here is the process we would recommend:
- Pick one repetitive task, such as weekly newsletter drafts or new product descriptions
- Write a standing compliance block and paste it into every prompt for that task
- Test the prompt on three real products or scenarios, not hypothetical ones
- Have one person review every output against your rules before it goes live
- Save the prompts that worked and note the ones that needed heavy editing
- Revisit the library every quarter, since both the tools and your regulations will change
The bottom line for Newport operators
AI prompts can save real time for a delivery service, but only when they are specific, tested against real inputs, and bounded by clear rules. The businesses that get value from them are not the ones with the cleverest prompts. They are the ones that treat each prompt like a small operating procedure, with an owner, a review step, and a record of what it is for. That discipline matters more in cannabis than in most industries, because a single careless sentence can create a compliance problem that takes weeks to unwind.
If you are just starting, keep the first project small. Pick one task, write the constraints first, and measure how much editing the output really needs. Once you trust the process, expanding it to other parts of the business becomes much less risky.