How to Write Better AI Image Briefs: A Product Manager’s Framework
· Mohammad Syed
Key takeaways
- Most AI image prompts fail because there’s no brief behind them. Treat the prompt as a tiny PRD for the image.
- A good brief has five parts: the job, the audience, one message, a structure and constraints.
- Decide what someone should understand, feel or do after three seconds of looking. If you can’t answer that, the image will try to say everything.
- Before you publish, check that it makes the point fast, the text is accurate, and it hasn’t invented stats, logos or product features.
Most AI image prompts fail before the image model sees them.
That sounds harsh, especially because the tools really are getting better. OpenAI’s latest ChatGPT Images release, for example, puts more emphasis on reference material, iterations and structured prompting. The capability is moving quickly.
But a better tool cannot rescue a brief that never existed.
“Make me a cool infographic about AI” is not a prompt with high standards. It is the visual equivalent of telling an engineer to “build something nice”, then acting surprised when the result is beige, busy and somehow includes a robot who did not need to be there.
The problem is rarely artistic taste. It is missing product thinking.
When I need an AI-generated visual to do real work, I treat the prompt as a tiny product brief. It needs a job, a user, one clear message, a structure and constraints. That is enough to move from “make something cool” to “make something useful”.
A strong image prompt is a small PRD for an image.
Why the model gives you generic work
An image model is very good at producing a plausible interpretation of the information you give it. When the information is broad, it has to make a lot of creative decisions for you.
That is how you get a beautiful but strategically empty image. Or a visual that contains five competing ideas, three fonts, an accidental brand identity and a level of futurism nobody asked for.
A useful image needs more than a style request. It needs a decision about the outcome.
Are you trying to explain a framework? Make a product feature easier to understand? Create a visual pause in a LinkedIn feed? Help a customer compare two choices? Give a blog article a mood without replacing the writing?
Those are different jobs. They deserve different briefs.
The five-part AI image brief
1. Define the job
Start with the work the image needs to do.
Avoid: “Create a visual about AI agents.”
Try: “Create a LinkedIn infographic that helps non-technical product leaders understand that an AI agent needs permission boundaries before it can take an external action.”
The second version gives the image a role in the reader’s understanding. It tells you what success should look like before you start choosing colours or illustration style.
A useful test is this: what should a person understand, feel or do after looking at this for three seconds?
If you cannot answer that, the visual is likely to try to say everything at once.
2. Name the audience
A visual for a product leader who already understands APIs is not the same as a visual for a small business owner who has heard the word “agent” thirty times and is quietly tired of it.
Describe who the image is for and what they already know. Include the level of detail they can comfortably absorb in the format you are using.
For a LinkedIn graphic, that normally means one idea, big labels and enough contrast to survive a phone screen. It does not mean shrinking a whole workshop onto one image and calling it an infographic.
3. Choose one message
This is the discipline most people skip.
You can have a complex topic. You still need one message.
For example:
Permission must follow context.
A prompt is the beginning of a brief.
A passing test is not production proof.
These are not the full article. They are the thoughts the reader should carry away.
Once that message is clear, you can remove anything that does not serve it. That is not dumbing the idea down. It is respecting attention.
4. Give the image a structure
A good prompt tells the model what is primary, what is secondary and what should disappear.
For an explainer, that might look like a left-to-right flow. For a comparison, it might be two columns. For a tutorial, it might be a five-step sequence. For a product announcement, it could be one strong product screen with a single outcome statement.
Structure is not decoration. It is the mechanism that helps someone understand the point without reading a paragraph first.
Here is a practical example:
Create a 4:5 portrait LinkedIn infographic. Put Untrusted content on the left, AI agent in the centre and Approval gate on the right. Show a red blocked route to Send / write / install and a green route to Ask a human. The one message is: Permission must follow context.
That is very different from asking for a “cool cyber AI graphic”. It tells the model the hierarchy, the content and the visual logic.
5. Set the constraints
Constraints are where you prevent the expensive surprises.
Include the format, aspect ratio, brand palette, exact labels, reference images, examples to follow and things the tool must not invent.
For my LinkedIn visuals, I keep the constraints simple:
1080 × 1350 px portrait format
dark navy or charcoal background
warm off-white type with muted green accents
one large message, readable on a phone
no robots, no glowing brains, no fake product screens
a small consistent footer
Your constraints will be different. The point is to make the non-negotiables explicit before the image model starts filling in the blanks.
A reusable prompt template
Copy this and replace the brackets:
Create a [image type] for [use case and audience].
Job: Help the viewer [understand / decide / feel / do].
Audience: [Who they are and what they already know].
One message: [The single takeaway].
Structure: [Describe the layout and hierarchy].
Style: [Palette, mood and visual language].
Exact labels: [Any wording that must appear].
Constraints: [Aspect ratio, brand requirements, references, elements to include or exclude].
Avoid: [Generic tropes, unwanted content, visual risks].
You do not need to write a novel. You need to remove the ambiguity that would change the outcome.
Example: turning a vague request into a useful visual
Here is the vague version:
Make me an infographic about evals.
Here is a more useful brief:
Create a 1080 × 1350 px portrait LinkedIn infographic for AI product managers. Job: explain that an AI demo becomes a product only when it is evaluated across accuracy, safety, latency and cost. Audience: PMs who understand product metrics but are new to AI systems. One message: Define good before you ship. Structure: a circular loop with four equally sized checkpoints around a central label, Evals. Style: dark navy, warm off-white, muted green, clean editorial type. Exact labels: Accuracy, Safety, Latency, Cost. Constraints: mobile-readable, one idea only, small footer Mo Syed | AI PM Field Guide. Avoid robots, code blocks, gradients and dense text.
The tool may still need iteration. That is normal. The difference is that every iteration now has something concrete to improve against.
Three checks before you publish
Before you post a generated visual, check three things.
First, does it make the intended point quickly? If the viewer needs your caption to decode the image, the visual is decorative rather than explanatory.
Second, is the text accurate and readable? Proper names, numbers, claims and safety labels are not places for a casual “close enough”. If a label matters, verify it. If it cannot render cleanly, simplify the image or use a format that can support the content properly.
Third, does it introduce anything you did not approve? A model can make a useful visual look more authoritative than it is. Watch for invented statistics, logos, UI elements, company claims and implied product capabilities.
The goal is not to make AI imagery feel perfect. The goal is to make it intentional.
The real lesson
The best image prompts are not magic words.
They are proof that someone has done the thinking before asking the tool to create.
That is why this is an AI product management lesson, not merely a design trick. Good prompts surface the same decisions good products need: purpose, audience, hierarchy, constraints and a clear definition of success.
Use the model for exploration.
Keep the judgment.
And please, for the love of all things visual, stop prompting vibes.