How to use AI for social media: a six-role system

Short answer: Use AI as six roles in a row: market researcher, brand director, social media creator, short-video producer, data analyst and WhatsApp customer service. Each role saves files that the next one reads. Write every prompt as role, background and task, and plan at least 30 days of content before you generate a single post.

Based on an evening class I taught in Hong Kong on 15 September 2026.

Sam Wong (黃力桐) is a Hong Kong corporate AI trainer and co-founder of Adaptig. 11,000+ people trained · 70+ organisations · 13 countries.

In September I taught an evening community class for people who sell and promote products themselves. Some were in the room, some on a livestream. We invented a product for the exercise, a daily wellness drink sachet, and took it to market with AI, one step at a time.

I should admit something first. I don't really like being on camera, and I don't really like posting on social media. What I told the room is that AI lets you hand over the part you are worst at and let it fill that gap. That only works if the AI is part of a system.

What are the six roles AI can play for social media?

In class, every step was a role you play with an AI tool. Each one builds on the step before it:

1. Market researcher. Know the market first: competitors, customers, and the gap.
2. Brand and marketing director. Set your look and your voice, so day 7 and day 30 still look like the same brand.
3. Social media creator. Platform guides, a content calendar, images and captions.
4. Short-video producer. Turn a written post into a spoken script, then into a short video.
5. Data analyst. Look at what worked, so you repeat the patterns instead of guessing.
6. WhatsApp customer service. An assistant that answers common questions from the same files.

Six roles, one frame, one folder.

Why should every AI prompt have a role, background and task?

Because one frame is easier to learn than twenty tricks. Every prompt I taught that evening opened the same way:

• Role: what you want the AI to be. "You are the copy director for [product]."
• Background: what it should refer to: the market research, the brand guide, the files you saved earlier.
• Task: what it does: a platform guide, a 30-day calendar, a caption.

The product name sits in brackets. Swap the bracket and the same prompt works for another product. As I said in class: "What we're learning is not memorising. We're learning the frame."

A lot of people tell me AI is getting stronger, so surely we can give it less. Actually, no. And you don't need ten or twenty prompting techniques. I learn one technique that works, and that's enough.

Why build a 30-day content system instead of one-off posts?

Because one or two posts and then silence does nothing. The pattern I see most: people generate some copy with AI for a day or two, and by day three they have stopped. That is not a system, and the effort is wasted.

What I teach instead:

• Ask the AI for a 30-day content calendar built from your research and brand files. In class it came back with six repeating formats: a scene, a personal story, product detail, customer questions, education, and an educational carousel.
• You don't need to learn all of marketing first. You need to know how to ask for the framework.
• Push the calendar into the diary you already use, such as Google Calendar or Outlook, so an assistant reminds you when to post.
• Save everything as files. Then every step has a record, instead of a chat that has forgotten what you said by next time.

"If you're serious, if you really want a system, build at least a 30-day system first."

How do you get AI images that look like your real product?

Give it pictures, not just words. At the image step I asked the whole room to run one prompt in Gemini, and told them first that most of them would not be able to generate it. Most were refused, because there were no reference pictures. If you push it to generate anyway, it makes random things up.

The fix was to change the inputs, not the prompt:

• Picture one: the packaging.
• Picture two: the product from another angle.
• Picture three: the logo.

The prompt says what each picture is. Run again, the image came back with our actual product design and the logo in place. No product to hand? Photograph a can of Coke or a cup at home and practise with that.

For the description itself I use a simple layer I call SOBER, and in class I only needed three letters of it. S is style, O is object, B is background. Think of a bowl of noodles: the noodles are the object, the soup base is the style, and the bowl you serve it in is the background.

How do you stop AI posts from sounding fake?

Check them, and put a person in the loop. Marketing and sales run on trust. If you use AI every day and don't check, sometimes it will simply be wrong. And customers who can tell you used AI feel you weren't sincere.

Here is how I put it to the room: I spent five seconds making a post, and I expect you to spend five minutes reading it and then pay me. That is an unequal relationship, and it is why more and more people dislike AI-generated content.

Two habits help:

• A second AI reviews the first. What did it miss, what did it say badly, what is wrong?
• A human shapes the result. Even with AI checking AI, there is human effort in the middle, and people can see it in the output.

Platforms differ too. When someone asked whether "sensitive to AI content" means a post gets blocked, my answer was: two things. Whether the platform pushes you, and whether the audience buys it. On Threads, if people think you're too fake they dislike you or ignore you, and the algorithm works against you. Facebook, not necessarily.

Can an AI WhatsApp assistant keep its answers short?

Yes, because you set it in writing. Someone in the room asked exactly this after I showed a WhatsApp assistant answering questions about our practice product. My answer: I set it, capped hard at four sentences. On WhatsApp, if you don't cap reply length, it pours out a lot at you, and I need it to sound like a real person.

Building a WhatsApp customer-service bot used to mean writing code, waiting for approval and hard-coding every reply. What I showed that evening ran on one computer and one folder: the product facts, guidelines, rules and pictures we had built earlier in the class. I did not teach the build that night. Done properly, it needs a three-hour class of its own.

Where should a team start?

With voice. Of everything in that class, voice is the one thing I asked people to go home and try. Most of the time I don't type when I use AI. Senior surgeons I trained earlier that year told me a big reason they disliked AI was typing on a small phone screen (AI for Surgeons case). I wrote up my own method in how to use ChatGPT at work without typing.

Then practise, and ask. Practising and asking questions is the fastest way to learn AI.

If you are planning this for a team, my page on corporate AI training in Hong Kong explains how I run sessions.

FAQ

What is the best way to use AI for social media content?

Treat it as a system rather than a simple post generator. In my class on 15 September 2026, we used six roles in sequence: market researcher, brand director, social media creator, short-video producer, data analyst, and WhatsApp customer service. Each role saves files that the next one reads. This keeps every post, image, and reply based on the same research and brand guidelines.

What should an AI prompt for social media include?

Every prompt should follow a frame of three parts: Role, Background, and Task. The Role defines what the AI should be (for example, the copy director for your product). The Background provides reference files like market research or brand guides. The Task specifies the output, such as a 30-day calendar. Keep the product name in brackets, and the same frame works across different products.

Why won't AI generate an image of my product?

Usually because it lacks reference pictures. In my September 2026 class, most people's Gemini refused the product-photo prompt until real images were added; pushed to generate anyway, it made things up. Provide photos of the packaging, the product from another angle, and the logo, and say in the prompt what each picture is. A simple phone photo of an object at home is enough to practise.

How long should an AI content plan be?

Build at least a 30-day content system. Creating one or two posts without a plan often leads to stopping by day three. A 30-day system uses your research and brand files to generate a calendar with repeating formats, such as scenes, personal stories, product details, customer questions and education. Add the plan to the calendar you already use, so you get reminded to post.

Can AI answer customers on WhatsApp?

Yes, by giving the AI one folder of product facts, guidelines, rules and pictures. To make it sound like a real person, set a hard limit on reply length; the assistant I showed in class was capped at four sentences. It runs on one computer and one folder, but building it properly is its own piece of work, about a three-hour class on its own.

Do AI-written posts hurt brand trust?

Unchecked AI content can harm trust, because marketing runs on sincerity and customers notice a post that took five seconds to make. Unchecked AI copy can also simply be wrong. To keep trust, use a second AI to review the first for gaps and errors, and have a human shape the final result. Some platforms, such as Threads, also react badly to content that looks fake.

Talk to me

If your team wants a social media system like this, message me on WhatsApp: Message Sam on WhatsApp

Sources

• Notes from my class recording, Hong Kong, 15 September 2026 (in Cantonese; quotes translated).
• AI for Surgeons case: /en/cases/ai-for-surgeons/