Last spring our operations lead spent most of a Monday trying to get an AI assistant to write a clean, compliant description for a new pre-roll pack. Every draft leaned too far into health claims, and every fix created a new problem. Somewhere in that afternoon, someone asked whether it made sense to look at chatgpt prompts for sale from people who had already solved similar problems, instead of rebuilding the same instructions by trial and error. That question turned into a small experiment, and the experiment changed how our team works with AI.
The blank-page problem in delivery operations
Cannabis delivery is a narrow business with a lot of words that have to be right. Product listings, age verification messages, delivery window texts, review replies, and driver notes all carry different rules. A friendly sentence in a menu can be a problem in a regulated market, and a short text to a customer can accidentally sound like advice. Most AI tools will happily produce fluent copy without knowing any of that.
Our early attempts followed a familiar pattern. Someone would type a vague request like “write a product description for this indica gummy,” get a paragraph back, and then spend twenty minutes removing phrases that were too strong. The output was not useless, but it was inconsistent. Two team members would ask for the same thing and get different tones, different lengths, and different levels of risk.
What makes a prompt actually work
A useful prompt is less like a question and more like a work order. After comparing our best and worst results, we settled on a few elements that consistently separated them:
- A defined role. The assistant is told it is writing for a licensed retailer’s menu, not a general audience or a wellness blog.
- Hard boundaries. A list of banned topics, such as medical benefits, dosing advice, or implied cures, stated plainly at the top.
- Input fields. Placeholders for product name, weight, strain category as listed on the label, and required disclaimer text, so nothing is guessed.
- An output format. A fixed length, a fixed order of sections, and a rule to return a short list of items the human reviewer should check.
- An example. One approved sample from our own menu, so the tone is anchored to something real.
When a prompt has these pieces, the results stop being a lottery. You still review everything, but the editing time drops and the drafts start to sound like they came from the same brand.
Where we use them day to day
We now maintain a short internal library with a few categories. None of these replace human judgment, but they remove a lot of repetitive drafting.
Product copy
Menu descriptions focus on flavor profile, packaging format, and serving size as printed on the label. The prompt explicitly forbids words that suggest therapeutic effects. Our compliance reviewer checks every line before it goes live, and the prompt returns a flagged-terms list so the reviewer knows where to look first.
Age and identity check messages
Customers sometimes get confused when a driver needs to verify ID at the door. We use a prompt that produces three versions of a message: a standard reminder sent the evening before, a same-day reminder, and a short line for the driver to read if a customer is unsure. The rule is simple: the message must state what is needed, what happens if it is not available, and nothing else.
Delivery window updates
Texts about delays used to be written in a rush by whoever was on dispatch. Now the prompt takes the original window, the new estimate, and the reason category from a dropdown, then returns a message under a set character count. This has made our updates more consistent and less apologetic in tone.
Review responses
Responding to reviews is where tone matters most. Our prompt asks for a response that thanks the customer, addresses the specific complaint in one sentence, and never discusses order details or customer identity. Replies to negative reviews are always reviewed by a manager before posting. To go deeper, explore The marketplace for AI prompts that actually work.
Driver and staff training
New drivers ask the same questions in their first week: what to do if a customer is intoxicated, how to handle a refused delivery, what paperwork to keep. We built a prompt that turns our policy document into a quiz with answer keys. The quiz is only as good as the policy it is built from, so the policy owner signs off on every version.
Guardrails we will not skip
Using outside prompts comes with real risks, and we set rules before anything touched a customer-facing channel.
- Human review on everything public. No AI output goes to a customer, a review site, or a social channel without a named person approving it.
- Local rules first. Each state and municipality has its own advertising and packaging requirements. A prompt written for one jurisdiction does not automatically fit another, so we adapt the constraints and have counsel check them.
- No medical or dosing guidance. Our assistants are never used to answer health questions. If a customer asks, the script directs them to a licensed professional.
- No customer data in prompts. Names, addresses, order histories, and ID details stay out of any prompt input. Placeholders are filled in only inside our own systems.
- Version control. Every prompt has an owner, a date, and a change log. When the rules change, the prompt changes, and old outputs are not reused.
How we evaluated prompts before buying
Buying a prompt sight unseen is a bad idea, and we learned that the hard way with one early purchase that produced confident-sounding copy full of claims we could not use. Our current process is more deliberate.
First, we ask for a sample output on a neutral input, not just the seller’s showcase example. Second, we run it against our own edge cases, including a product with a long name, an order with a missing field, and a request that tries to push toward medical language. Third, we check the license terms: can we edit the prompt, use it commercially, and keep using it if the seller stops updating it? Fourth, we look at whether the listing explains what the prompt does not do. A prompt that is honest about its limits is usually easier to trust.
We also compare the prompt’s output against our own style guide. If a prompt produces copy that would never pass our review, the price does not matter. We have walked away from listings that looked impressive on the surface but failed on the second test.
What we would tell another delivery operator
Start with one workflow, not ten. Pick the message your team rewrites most often, build a prompt around it with clear boundaries and a fixed format, and measure how much editing time it saves over two or three weeks. If the savings are real, expand. If they are not, the problem is probably the input data or the policy behind it, not the AI.
Keep your compliance lead involved from the beginning. The most useful thing they can do is write the list of forbidden phrases and required disclaimers, because that list becomes the backbone of every prompt. And remember that a prompt is a tool for drafting, not a replacement for judgment. The best result we get is a draft that a trained person can approve in under a minute, with nothing in it that needs to be explained to a regulator.
Our team still writes a lot by hand. But the blank-page afternoon is gone, and the drafts we send out now look and sound like they came from one careful, consistent voice. For a delivery business that operates under strict rules, that consistency is worth more than any single clever sentence.
Leave a Reply