Our dispatch and content team is small, and every week we hit the same wall: a seasonal menu banner is due, a neighborhood post needs writing, and the blank page keeps winning. When we first experimented with visual content, we searched for ready-made midjourney prompts to see whether a tested instruction could beat our clumsy first attempts. It could, but only after we learned what separates a prompt that works from one that merely looks impressive in a listing.
What "actually works" means in practice
A prompt that works produces a usable result on the first or second try, with predictable quality across different subjects. Most prompts fail for boring reasons: they name no composition, give no lighting direction, and leave the output format to chance. The ones we kept shared a few traits:
- A clear role or style reference, such as a travel photographer working in golden-hour light.
- A specific subject with concrete details, like the rocky arch at Land’s End or the marina at dusk.
- Explicit constraints on color palette, framing, and what to leave out.
- A stated output format, so the result fits a banner, a story frame, or a menu card without cropping.
We also started logging every prompt we ran, the version of the tool, and a one-line note on whether the output was usable. After a few weeks, that log mattered more than any single prompt.
Guardrails for a cannabis business
Cannabis is the part of our work that needs the most care, and no image generator knows our local obligations. Mexico’s cannabis rules are still developing, and what a business may advertise, and where, depends on current law and on the specific activity involved. We are not lawyers, so we treat the following as internal policy rather than legal advice, and we run anything customer-facing past qualified counsel before it goes live.
- We never generate images that depict consumption, intoxication, or people appearing to be minors.
- We avoid health or therapeutic claims in any generated copy, even when a prompt suggests them.
- We do not use prompts that imitate packaging from other brands.
- Every generated image is reviewed by a human before publication, because models sometimes add details nobody asked for, such as stray text or incorrect signage.
The safest visual we have produced so far is a non-cannabis scene: a delivery scooter crossing a quiet street in San José del Cabo at sunrise, with the brand color used only on the bag. It photographs well, it is unambiguous, and it avoids the compliance questions entirely.
A prompt structure we reuse
Rather than collecting dozens of disconnected instructions, we settled on one template and vary only the blanks:
Role: a commercial photographer specializing in coastal Mexico. Subject: a single delivery scooter on a tree-lined street in San José del Cabo, early morning. Lighting: soft low sun from the east, warm but not orange. Composition: subject in the lower third, open sky above for text overlay. Constraints: no people in frame, no readable logos, no cannabis products visible. Output: 4:5 vertical for social, with clean edges.
The same skeleton works for menu headers, event posters, and email banners. When a result misses, we change one variable at a time so we can tell which edit helped. To go deeper, explore The marketplace for AI prompts that actually work.
Evaluating a prompt marketplace before you buy
If you are weighing whether to build your own library or buy one, a curated collection such as the reviewed prompt library at PromptMart is one option to compare against your own testing. Whatever source you choose, check these points first:
- Sample outputs that match the prompt, not just a cover image.
- Clear notes on which tool and version the prompt was tested on, since results drift between releases.
- Plain license terms covering commercial use of the generated images.
- A refund or replacement policy for prompts that do not perform as described.
- Categories that fit your niche, so you are not paying for prompts you will never run.
Price matters less than verification. A cheap prompt that you must rewrite is more expensive than a well-documented one that works the first time.
Protecting customer data in your prompts
One habit we would urge on any delivery business: never paste customer information into an AI tool. That includes names, addresses, order histories, and phone numbers, even when you want help writing a reply. We keep a rule that prompts describe the situation in generic terms, such as a customer who asked about delivery windows, and our staff fills in the personal details afterward inside our own systems. It takes a few extra seconds and removes a whole category of privacy risk.
How we measure whether it is working
We do not chase precise performance numbers we cannot verify. Instead, we track simple operational questions: how long a banner takes from brief to approved draft, how many generations it takes to get one usable image, and whether the team actually reuses a prompt a month later. Prompts that pass all three tests stay in our shared library. Prompts that do not get retired, regardless of how good their sample looked.
The takeaway
AI prompts are a tool, not a strategy. They help a small team move faster on the routine parts of marketing, but the judgment about what is lawful, accurate, and appropriate for your customers stays with you. Test before you trust, keep a log, respect the rules that apply to your product, and keep people in the review loop. For a Los Cabos delivery brand, that combination has saved us hours each week without putting the business at risk.
Leave a Reply