Our dispatch lead once spent a full Saturday morning asking an AI tool to rewrite our product descriptions, and the results were so generic that nobody on the team could use them. Searching for chatgpt prompts for sale was the first thing that got us out of that rut, because it showed us that other businesses had already figured out which instructions produce usable output and which ones produce filler. We are a cannabis delivery service in Newport, and we are not a software company, so this article is an honest account of what we bought, what we built ourselves, and what we still do by hand.
Why a delivery service bothers with prompts at all
Delivery is a logistics business wearing a retail costume. Every day we write the same kinds of text: order confirmations, “your driver is two stops away” messages, out-of-stock notices, a weekly staff briefing, and product copy that has to stay inside tight advertising rules. Each of those tasks is small, but together they eat hours. When an AI model can draft a first version in seconds, the time saved is real, provided the draft is close enough to send after a light edit.
The problem is that most people, including us at the start, type something like “write a product description for a pre-roll” and get back paragraphs full of superlatives that would get any licensed operator in trouble. A prompt that works has to encode our constraints, our voice, and our format before the model writes a single word.
What separates a working prompt from a wasted one
After a few months of testing, we noticed that the prompts that held up shared the same structure. They were not clever. They were specific.
State the role and the reader
A prompt that begins with “You are writing for a customer who has ordered from us before and wants a quick status update, not a sales pitch” produces a different result from one that just says “write a message.” The reader is the single biggest variable, and naming them narrows the output more than any other single change.
Give the boundaries explicitly
Models do not know what you are not allowed to say unless you tell them. We now include a short list of banned phrasings in every prompt that touches product language: no health or therapeutic claims, no implied effects, no pricing promises, and no language aimed at anyone under the legal age. Writing those rules once, at the top, saved us from editing the same mistakes over and over.
Specify the output shape
If you need three subject-line options under 50 characters, say so. If you need a driver briefing with a bulleted list of addresses in route order, say that too. Vague format requests produce vague text. Precise format requests produce text you can paste into a template without rewriting it.
Include one real example
The single most useful addition we made was a sample of a good message we had already sent to a customer, marked as a model answer. Pasting one approved example into the prompt improved tone more than three paragraphs of adjectives describing the tone we wanted.
Where our team actually uses them
We do not let AI talk to customers unsupervised, and we do not let it touch order data. Here is how the work is divided in practice: To go deeper, explore The marketplace for AI prompts that actually work.
- Order status templates. We generate variations of the “out for delivery” and “delayed due to weather” messages so they do not all read identically. A staff member reviews each batch before it goes into the system.
- Driver shift briefings. Drivers get a short summary of the day’s zones, known parking problems, and reminders about verification steps. The model turns our messy notes into a clean one-page sheet.
- FAQ drafts. Questions like “what happens if no one is home” or “how do I update my delivery address” get a first draft, which our operations manager then checks against our written policy.
- Internal training scenarios. New dispatchers practice with made-up but realistic situations, such as a customer who insists the driver should skip ID checks. The model plays the customer, and the trainer scores the response.
Notice what is missing from that list. We never ask a model to invent product claims, to recommend strains for a medical condition, or to write anything that goes out without a human reading it first.
Compliance is not optional, and prompts can make it worse
A poorly written prompt can produce confident, fluent text that breaks the rules without anyone noticing, because the language sounds professional. That is the real risk. Before we publish or send anything, we check it against a short internal checklist:
- Does the text make any health, medical, or effect claim, even indirectly?
- Does it mention price, discounts, or promotions that we have not approved in writing?
- Does it reach anyone who could be a minor, including through imagery descriptions?
- Would it be fine if a regulator read it next to our license?
If the answer to any of those is uncertain, the text does not go out. This checklist is more important to us than any prompt library, and no prompt, however well built, replaces it.
Buying prompts without getting burned
If you are considering paying for prompts, whether for your own business or for a client, treat them the way you would treat any outsourced template. A few questions help separate useful material from recycled filler:
- Is the prompt specific to a task? “Write marketing copy” is not a prompt. “Write a 60-word delivery window reminder for customers who ordered after 8 p.m.” is.
- Does the seller explain the inputs? Good prompts tell you what information to supply and what the model should do if that information is missing.
- Can you test it yourself? Run the prompt on your own data before relying on it. If the output needs heavy rewriting every time, the prompt is not doing its job.
- Does it respect your industry’s rules? A prompt written for a general retailer may encourage claims that are illegal for a regulated product. Adapt it or discard it.
We have also learned to keep our own prompt log. Each entry records the task, the prompt, a sample output, and a note on what we changed. Over time that log becomes more valuable than anything we purchased, because it reflects our actual customers, our actual zones, and our actual rules.
What we would tell another local operator
Start with one repetitive task that has low risk, such as internal shift notes or draft FAQ answers. Build the prompt with a role, boundaries, a format, and one approved example. Have a person review every output for the first few weeks, and keep notes on where the model drifts. Only expand to customer-facing text once the review process is routine.
AI prompts are not magic, and they will not fix a weak operation. What they do well is remove the blank-page problem from routine writing, leaving your staff more time for the parts of the job that require judgment, like handling a customer who is upset about a late order or a driver who needs a route change in the rain. Used that way, a well-built prompt is simply another tool in the dispatch office, and it earns its place only when it makes the next delivery run a little smoother.
Leave a Reply