The Marketplace for AI Prompts That Actually Work: Notes From a Honolulu Delivery Team

When our team first started experimenting with AI image generators for menu graphics, social posts, and educational cards, we assumed the hard part would be the software. It wasn’t. The hard part was words. Typing a loose idea into a generator rarely produced anything usable, so a few of us began searching for ready-made midjourney prompts to see what tested wording looked like. That search taught us more about prompt writing than any tutorial did.

Why most prompts fail for local businesses

Most prompts floating around online are written for novelty, not for a specific business with a specific audience. They describe a mood but skip the details that a shop actually needs. In our experience, the common problems were:

• Vague subjects, such as “a nice cannabis shop” with no clear scene or product type.
• No framing instructions, so the generator picked a random camera angle and crop.
• Missing negative constraints, which left out text, logos, or unwanted objects that then had to be removed by hand.
• Style words that conflicted with each other, like “photorealistic watercolor cartoon,” which produced muddy results.
• No aspect ratio, so a graphic meant for an Instagram square ended up as a tall banner.

A prompt that works is less like a poem and more like a short creative brief. It tells the tool what to show, where, how it should be lit, and what must not appear.

What a prompt that actually works looks like

After testing dozens of variations, we settled on a simple structure that holds up across different image models:

• Subject: one clear focal point, described in plain nouns.
• Setting: a specific location or background, such as a low-lit living room or a sunrise coastline.
• Lighting and time of day: soft morning light, warm evening glow, or overcast daylight.
• Composition: wide shot, close crop, empty space on the left for headline text.
• Style: one primary style only, such as clean flat illustration or natural photography.
• Constraints: no text, no logos, no people, no packaging labels, whatever your project requires.
• Aspect ratio: stated explicitly so the output fits the platform.

Once we adopted that structure, our rejection rate dropped noticeably. The change was not about finding magic words. It was about removing ambiguity so the tool had fewer decisions to guess at.

Where a prompt marketplace fits into the workflow

Writing a strong prompt from scratch is a skill that takes time, and not every member of a small team has the bandwidth to develop it. Some of us found it useful to study prompts that other creators had already refined and sold through a marketplace where tested prompts are listed with example outputs. The value was not in copying a single string word for word. It was in seeing how experienced writers handled lighting, framing, and constraints, then adapting those ideas to our own brand and our own compliance limits.

Before buying or adapting any prompt, check three things: whether the example outputs match the style you need, whether the description explains the model it was tested on, and whether the seller describes how it was revised. A prompt with no testing notes is a guess, however polished it sounds.

Compliance comes before creativity

This is the section we wish someone had written for us at the start. Cannabis is a regulated product, and marketing visuals carry legal risk even when they look harmless. AI tools make it easy to produce a lot of content quickly, which means it is also easy to produce a lot of problematic content quickly. Our working rules are below. They are not legal advice, and you should confirm current requirements with a qualified attorney and the relevant Hawaii agencies before publishing anything.

• Never generate people who appear to be minors, or any image that could read as youth-oriented, including cartoon characters, bright candy-style packaging, or playful mascots.
• Avoid images that show or imply consumption, intoxication, or smoking in a way that could appeal to non-adults.
• Do not include health or medical claims in generated images or captions, even implied ones such as a glowing “cure” motif.
• Keep product depictions accurate. If you sell a specific product, use your real packaging photos rather than AI reinventions of it.
• Store your prompts, dates, and final outputs together so you can show what was published and when.
• Run every final graphic past the same review checklist you use for printed materials.

We also made a rule that AI images are never the sole source of product imagery. Generated scenes serve as backgrounds, texture, and mood. Real product photos, verified by staff, carry the actual information customers rely on.

Building Honolulu-specific visuals responsibly

Our customers are local, and they respond to places they recognize. That gave us an opportunity to use AI for scenery rather than for anything that could cross regulatory lines. A prompt describing early morning light over the Honolulu coastline, empty sand, and a clean, simple sky gave us a calm header image for our delivery-zone announcements. We avoided stereotypical imagery and never used scenes that suggested the product was being used outdoors or in public.

For seasonal updates, we keep a short list of location-based settings: a quiet garden at dusk, a rainy windward road, a bright market table with fruit and no branded items. Each one is paired with a fixed style and aspect ratio so the set looks consistent across posts.

Testing and keeping a personal prompt library

The most useful habit we built was a simple spreadsheet. Each row records the prompt text, the model and version, the seed if available, the aspect ratio, a short rating, and what needed fixing. When a prompt produces a strong result, we note exactly what made it work. When it fails, we record the failure mode, such as a warped hand or unwanted text, so the next revision avoids it.

After a few months, that spreadsheet became more valuable than any single prompt. It showed us which style words were reliable, which constraints prevented the most cleanup, and which settings consistently gave us usable headers without touching an editor.

A short checklist before you publish

• Does the prompt state one clear subject and one setting?
• Have you specified lighting, framing, and aspect ratio?
• Have you listed what must not appear, including text and logos?
• Does the image avoid anything youth-oriented or suggestive of consumption?
• Is any product shown an accurate photo, not a generated one?
• Is the final version logged with its prompt and date?
• Has someone other than the person who made it reviewed it?

The takeaway

AI image tools are useful for a small cannabis delivery business, but only when the prompts are specific and the compliance rules come first. A good prompt marketplace can shorten the learning curve, and a personal log can turn each experiment into a lasting asset. Start with one clear brief, test it against your own rules, and keep your real product imagery at the center of everything you publish.