Running a cannabis delivery service in Portland means writing a lot of words every day: menu descriptions, replies to customers asking about delivery windows, driver handoff notes, and social posts that have to stay inside the lines. Many of us have tried AI writing tools to speed this up, and the early results were uneven. The fix was not switching tools. It was writing better instructions. Our team began looking at prompt engineering templates as a starting point, then adapting each one to our own product sheet, our brand voice, and the rules we have to follow.
Why generic prompts fail in a regulated niche
A vague request like ‘write a description for our new gummies’ usually produces copy that sounds like a wellness brand. It may promise relaxation, mention sleep, or describe effects in ways that could be read as health claims. In cannabis, that kind of language creates real risk for a business, and it can also reach people who should not be targeted at all.
The problem is not that AI is careless. It simply fills gaps with whatever is most common in its training material. If your prompt leaves gaps around facts, limits, and audience, the output will fill them with generic marketing habits. A working prompt closes those gaps before the first word is generated.
What a working prompt looks like for us
After several rounds of testing, we settled on a structure that every prompt in our shared document follows. It has four parts.
1. Give it the facts, not the vibe
We paste the exact product details from our menu sheet: product name, form, package size, ingredients as listed on the label, and the lab-tested potency values printed on the packaging. The model is told to use only these facts. If a detail is missing, it should say so instead of guessing.
2. State the boundaries first
Before any task instruction, the prompt lists what the output must not do. It must not make health or medical claims. It must not describe effects. It must not use imagery aimed at young people, such as cartoons or candy-style language. It must not suggest that cannabis is safe for anyone under 21. Putting these rules at the top keeps them from getting buried under the creative request.
3. Define the output format
We specify length, tone, and structure. For a menu description, that means two sentences, plain language, no exclamation points, and a closing line about adult verification. For a text reply, it means one short paragraph and a clear next step. Vague formats produce vague copy that needs heavy editing.
4. Ask for a self-check
Every prompt ends with a request for the model to list any claim it made and point to the fact that supports it. This is not a perfect safeguard, but it makes review faster. Anything without a source gets cut.
Where AI prompts earn their place in delivery operations
Not every task needs AI, and we have stopped using it for a few things entirely. But several routine jobs have become much faster with well-built prompts:
- Order-status replies that explain delays without revealing customer details
- Menu descriptions for new arrivals, using only label information
- Driver handoff notes that summarize special instructions for a drop-off
- FAQ answers about delivery zones, hours, and ID requirements
- Onboarding checklists for new team members learning our verification steps
- Weekly summaries of inventory changes for the operations meeting
In each case the time saved comes from not starting from a blank page. The human still checks the final result. The prompt just gets the first draft closer to what we would have written ourselves. To go deeper, explore The marketplace for AI prompts that actually work.
Building a shared library instead of starting over
Early on, each team member wrote their own prompts and kept them in private notes. Quality varied widely, and nobody knew which version had been approved. We moved to a shared document and later started comparing options in a curated prompt library for business operations, which gave us a few tested drafting frameworks to adapt rather than inventing everything from scratch. We still rewrite most of them to match our local details, but starting from a structure that already handles boundaries and formatting saves a lot of trial and error.
Guardrails we set before anyone pastes anything
A tool is only as safe as the habits around it. These are the rules our team agreed on before using AI for any customer-facing work:
- Never paste customer names, addresses, phone numbers, or order histories into an AI tool.
- Every customer-facing message gets a human review before it is sent.
- All marketing copy is checked against current advertising rules from the Oregon Liquor and Cannabis Commission. Because rules change, we assign one person to confirm the latest guidance each quarter.
- No copy describes effects, dosage for a purpose, or use for any medical condition.
- Approved prompts are logged with the date and the reviewer’s initials, so we know which version produced which output.
- Age verification and ID checks at delivery are never delegated to any automated process. These stay with trained drivers.
A simple testing routine
Before a prompt goes into the shared library, we run it three times with the same inputs. We compare the outputs against a short checklist: Is every fact traceable to the product sheet? Is the tone calm and factual? Does it avoid anything that could appeal to minors? Does it stay within the requested length? A prompt that passes all four checks on all three runs is approved. Anything else goes back for revision.
This sounds slow, but it takes about fifteen minutes per prompt, and it prevents the much longer process of pulling a bad message after it has already gone out.
What we still do by hand
Some work will never be handed to a prompt. Driver safety decisions, ID checks at the door, responses to complaints involving health or legal concerns, and any question about what a customer is allowed to purchase stay with people who have the training and accountability for those calls. AI helps with the paperwork around those tasks, not the tasks themselves.
Getting started this week
If your delivery business is thinking about AI, start small. Pick one repetitive task, such as order-status replies or FAQ drafts. Write a prompt that includes your verified facts, your boundaries, and your output format. Run it three times, review the results against your compliance rules, and log the version you approve. Once that process works for one task, repeat it for the next.
The goal is not to publish more words faster. It is to spend less time fixing avoidable mistakes so your team can focus on what matters most to customers: accurate information, reliable delivery, and a careful approach to a product that still carries real restrictions.









