Running a cannabis delivery service in San Francisco means writing a lot of copy: menu descriptions, weekly specials, delivery-zone updates, email blasts, and social posts that have to stay inside state advertising rules. Many owners try an AI writing tool once, get a generic paragraph back, and give up. The problem is rarely the tool. It is usually the prompt. If you are also experimenting with image generation for branded visuals, you may have already searched for midjourney prompts and noticed how much the quality of a short instruction changes the output. The same principle applies to text work, and it is the foundation of this guide.
Why Most AI Copy for Cannabis Falls Flat
A typical first attempt looks like this: “Write a product description for a indica gummy.” The result tends to be vague, overly enthusiastic, and sometimes full of claims that a compliance reviewer would flag immediately. Shoppers on a delivery app skim quickly. They want to know the product type, the format, the dose per piece, the ingredients, and what the package actually contains. A generic paragraph gives them none of that.
A usable prompt does three things. It assigns a role, it supplies the facts you already have, and it sets boundaries on what the model must not say. For a delivery business, those boundaries matter more than creativity. You cannot let a tool invent effects, health benefits, or potency numbers that do not match your lab results.
Build a Prompt Template Before You Build a Prompt Library
Start with one template and test it against real products from your current catalog. A workable structure for a product description looks like this:
- Role: You are a product copywriter for a licensed cannabis delivery service in San Francisco.
- Facts: Paste the exact product name, format, package count, THC and CBD values from the certificate of analysis, flavor notes from the producer, and the ingredient list.
- Length and format: Two sentences for the menu card, one bullet list of three features, and a 140-character version for the app listing.
- Restrictions: Do not mention health outcomes, medical conditions, or effects. Do not use the words cure, treat, or heal. Do not reference minors, youth-oriented imagery, or prices.
- Tone: Plain, specific, and calm. Avoid superlatives.
Once the template produces consistent output for ten or fifteen products, you can save it and reuse it. The value is not in any single clever sentence. It is in the fact that every new product description follows the same disciplined path.
Prompts for the Jobs Delivery Teams Actually Do
Menu and Catalog Descriptions
Feed the model one product at a time, using the template above. Ask for output in a fixed format so your team can paste it directly into your point-of-sale or menu platform. If a description comes back with information you did not supply, treat that as a failure of the prompt and tighten the restrictions rather than editing the output by hand every time.
Weekly Specials and Email Updates
Specials are where teams tend to drift into promotional language that may not be allowed. A safer prompt gives the model the product, the date range, and a required disclaimer line, then asks for a short email body. Your compliance contact should approve the disclaimer wording once, and the prompt should always include it verbatim. Do not ask the model to write the disclaimer itself.
Delivery Zone and Schedule Notices
Customers in neighborhoods from the Mission to the Sunset need to know when cutoffs change, when holiday hours apply, or when a zone is temporarily unavailable. These messages are operational, so they should be short and literal. A prompt that says “Write a three-sentence notice. State the new hours exactly as given. Do not add any other information” will outperform a broad request every time.
Social Posts That Stay Within Bounds
Social platforms add another layer of restrictions, and rules differ by channel. Build a separate prompt for each platform, with its character limits and its own list of banned claims. Ask for three options, then have a human pick one. Never publish the first output without review.
Testing and Scoring Your Prompts
A prompt is only good if its output passes review consistently. Create a simple scoring sheet with four columns: factual accuracy against your source data, compliance with restrictions, readability for a shopper scanning on a phone, and how much editing was needed. Run each prompt on five real products and score the results. Any prompt that scores poorly on accuracy or compliance goes back to the drawing board before it reaches your menu.
Keep a version history. When you change a restriction or add a new product format, note the date and the reason. This habit becomes valuable during audits or when a new team member needs to understand why the copy looks the way it does.
Where to Find and Share Prompts That Work
Many operators assume they must invent every prompt from scratch. In practice, the faster route is to borrow proven structures and adapt them to your rules. Marketplaces that collect tested prompts, with examples and buyer feedback, can save hours of trial and error. Look for listings that show sample inputs and outputs so you can judge quality before paying, and check whether the seller explains how the prompt handles restricted content. If you want a starting point for structured prompt libraries, the PromptMart marketplace is one place to browse prompts across writing and image categories. Treat anything you buy as a draft: adapt it to your template, run it through your scoring sheet, and keep your own compliance language in control.
Keeping a Human in the Loop
AI tools do not know your license terms, your local rules, or the difference between a product that is in stock and one that is about to be discontinued. Someone on your team should own every piece of copy before it goes live. Set a simple workflow: the prompt generates, a second person checks against the source sheet, and a named manager approves. This takes minutes per item and prevents the errors that cause the most damage.
Also decide what the tool will never write. Most delivery businesses should keep AI away from anything involving health, dosing advice for specific conditions, age-related messaging, and customer complaints. Those topics need human judgment and, in many cases, legal review.
A Practical Starting Plan for This Month
- Choose ten products from your current catalog that represent your range: flower, pre-rolls, edibles, and one or two accessories.
- Write one product description template using the structure above, and confirm the restriction list with your compliance advisor.
- Generate descriptions for all ten, score them, and revise the prompt until at least eight pass without major edits.
- Create a separate prompt for your weekly email and one for delivery-zone notices.
- Save everything in a shared document with version dates, and assign one person to maintain it.
The Bottom Line for San Francisco Delivery Operators
AI can save real time on the repetitive writing that fills a cannabis delivery business week after week. The gains come from disciplined prompts, clear restrictions, consistent testing, and a human who signs off on every word. Start small, measure accuracy, and expand only when the process holds up. Customers in San Francisco value clear information and reliable service, and copy that reflects both is good for your brand and for your compliance record.
Leave a Reply