Writing Better AI Image Prompts for Cannabis Delivery Marketing in Baltimore

Every few weeks, someone on our team asks whether AI image tools can make our delivery menus and social posts look professional without a photographer on retainer. The short answer is yes, but only when the instructions behind the image are specific. Many of us started by searching for midjourney prompts that other people had shared online, and we quickly learned that a vague phrase like “cool weed aesthetic” produces images that fit no brand, no neighborhood, and no set of advertising rules. This article covers how we approach AI prompts for a Baltimore cannabis delivery business, what we keep out of our creative process entirely, and how we turn a rough idea into something we can publish.

Why prompt quality matters more for regulated businesses

A general consumer brand can often get away with a loosely worded image request and a quick edit. A licensed cannabis business has less room for error. Imagery can attract attention from regulators, platforms, and customers who are not yet of age, so a prompt that produces the wrong visual can create real problems. Precision is not just about aesthetics. It is a way of controlling risk.

For our delivery service, the most useful prompts describe the scene in plain terms: who or what appears, where it is, what the light looks like, and what must not appear. When we leave any of those elements open, the tool fills the gap with its own assumptions, and those assumptions are often the wrong ones.

Compliance comes before creativity

Before anyone opens an image tool, we check the request against our compliance rules. Maryland rules on cannabis advertising change, so we review the current guidance from the Maryland Cannabis Administration rather than relying on memory or on what another business did last year. Our internal checklist currently covers the following points:

• No cartoon characters, animals, toys, candy-like shapes, or anything else that could appeal to children.
• No people who look under 21, and no scenes that suggest a school, playground, or youth activity.
• No health, medical, or therapeutic claims in the image or in any text placed over it.
• No depictions of consumption, intoxication, or driving.
• No product imagery that misrepresents potency, packaging, or what is actually sold by our license.

We write these restrictions directly into the prompt as negative instructions, and we also review every output by hand. An AI tool does not know which license you hold or which county you operate in, so the responsibility stays with the business.

A prompt structure that works for us

After many rounds of trial and error, our prompts follow a consistent structure. We do not claim that this structure is magic, but it has made our drafts more predictable and easier to review.

1. Subject and setting

Start with one clear subject and one clear place. For example, instead of “a happy customer with product,” we might describe an adult in everyday clothing opening a plain paper delivery bag at a front door in a row house neighborhood during late afternoon. The more concrete the setting, the less the tool invents.

2. Local texture

Baltimore has a distinct visual identity: brick rowhouses, marble front steps, harbor light, and neighborhood murals in places like Hampden and Fells Point. Naming these details helps keep our imagery grounded in the city rather than in a generic urban stock photo. We only reference real places when the image does not imply any endorsement by a venue or business that has not agreed to it.

3. Lighting and camera language

Photography terms give the tool clearer direction than adjectives like “beautiful” or “modern.” Phrases such as soft window light, shallow depth of field, overcast daylight, or eye-level framing have produced more consistent results for us than mood words alone.

4. Explicit exclusions

We end nearly every prompt with a list of what to avoid: no text overlays, no logos, no visible brand names, no people who look young, no cannabis leaves as decoration. Exclusions matter because many tools default to adding these elements.

Turning a rough idea into a publishable image

Our workflow has four stages, and skipping any of them has caused us problems in the past.

1. Brief. One staff member writes the goal in a sentence: what the image is for, which channel it will run on, and which promotion or message it supports.
2. Draft prompt. We use the structure above and keep a short log of what we changed and why.
3. Compliance review. A second person checks the output against the current checklist. This person is allowed to reject an image even if it looks excellent.
4. Final text. Any headline or caption is written separately and reviewed for accuracy, pricing, and required disclaimers. We never let the image model generate text that will appear on a published ad.

Separating these steps has a practical benefit. When a platform questions one of our posts, we can show the brief, the prompt history, and the review notes instead of reconstructing them from memory.

Where outside prompt libraries fit in

We do not think a prompt library replaces judgment, but it can save time at the beginning. When we want to see how other people phrase a request for a specific style, a curated collection is faster than starting from a blank page. A marketplace such as PromptMart lists prompts organized by use case, which can be a sensible starting point if you treat every listing as a draft to adapt rather than a finished instruction set. Always rewrite the parts that touch on people, products, or claims so they match your own license and your own rules.

What we avoid is copying a prompt that was written for a different industry. A prompt built for sneaker advertising or travel photography will carry assumptions about lifestyle, age, and product that may not fit a regulated business at all.

Common mistakes we have made

• Trusting the first output. The first image is almost always the most generic. Refining the prompt usually improves the result more than generating dozens of random variations.
• Forgetting the platform. Images that work on a website banner may be rejected by a social network with stricter cannabis policies. Check each channel separately.
• Letting the tool add text. Many models render slogans and labels that look official but are inaccurate or misleading. We remove all generated text.
• Skipping the log. Without notes, you cannot explain why an image was approved or what changed between versions.
• Assuming a fast turnaround means a safe image. Speed is useful only after the compliance step is complete.

A short checklist before you publish

Use this list every time, even for a small social post:

• Does the image match the current Maryland cannabis advertising guidance we follow?
• Is every person clearly an adult, and is no one shown consuming cannabis?
• Is there any cartoon, toy, or child-appealing element?
• Does any text, including captions, avoid health claims and accurate pricing issues?
• Has a second person reviewed the image and the prompt log?
• Is the image appropriate for the specific channel where it will run?

Final thoughts

AI image tools can help a small Baltimore delivery team produce consistent visuals without a large creative budget, but they do not understand regulation, neighborhood sensitivity, or the difference between a safe image and a risky one. The advantage comes from clear prompts, documented reviews, and a willingness to reject output that looks good but breaks the rules. If you build those habits first, the creative part becomes much easier, and the business stays on solid ground while you experiment.