I’ve always leaned on images when I present. A slide with a strong visual does a lot of work, especially with a mixed audience. This time I was preparing a community Saturday Breakfast Forum talk titled “AI at the Crossroads: What We Gain, What We Risk, and What We Should Demand.” The crowd would be full of thoughtful professionals, curious and skeptical, but not necessarily steeped in AI jargon. I wanted the slides to carry the structure of the talk, and I wanted the visuals to do real explanatory work, not just serve as decoration.
The problem is that stock photos are a weird kind of time sink. You can spend 30 minutes searching for “the perfect image” and still end up settling for something that’s close, but not quite right. Wrong mood. Wrong metaphor. Wrong vibe. And even when you find a good one, it rarely matches the next one, so the deck starts to look like a patchwork quilt.
This time I tried something new. I used the image creation feature inside Apple Keynote’s Creator Studio to generate the illustrations I needed, right where I was building the slides. Then I used ChatGPT as a creative partner to help me find good metaphors and write prompts that the image tool could actually follow. It changed my whole workflow.
Why this is worth doing
Stock photos are great when you need something literal. They’re terrible when you need a visual metaphor. How do you find a stock photo for “AI as a confident liar,” or “customized scams at scale,” or “data becoming centralized power?” You usually can’t. So you either compromise on the message or you add more words to your slide to explain what your image can’t. Generating images inside Keynote flips that. You can design a visual that matches the idea.
One of the underrated benefits is consistency. If you lock a style and stick to it, the whole deck looks like it belongs together. It feels intentional. It feels designed. That alone can make a presentation feel more polished than a deck assembled from whatever images you could find at the last minute.
It also lets you design for your actual room. I knew I needed slides that would read quickly, from a distance, in real-world lighting. That meant strong contrast, bright readable colors, simple shapes, and minimal clutter. Stock photos do not care about your projector. Generated visuals can.
The workflow I used
Before I generated anything, I built the overall look of my slides in ChatGPT. That included a consistent color palette, typography rules, layout logic, and a repeating structure for benefits and risks. Once I had that locked, the images had a clear job. They weren’t decoration, they were part of a system.
Next I locked a “house style” for images. I chose a consistent 3D illustration look, not photorealistic and not cartoonish. Then I kept reusing the same style language in prompts, modern 3D illustration, clean shapes, high contrast, minimal backgrounds, no text in the image, no logos, big readable elements. That style language became my visual glue.
House style:
3D animated illustration, modern cinematic 3D style, high quality 3D rendering, bright studio lighting, very high contrast, saturated but harmonious colors, matte or satin materials, large readable shapes, minimal clutter, crisp edges, no logos, no text.
Then add: a clear description of the subject and layout, including the main objects, what they are doing, what should be emphasized, what should be left out, and any needed negative space for slide text.
Then I stopped jumping straight to prompts. Instead, I’d brainstorm metaphors first. What is the idea? What is the simplest object metaphor that communicates it? What can be shown without faces? What will still read on a projector? ChatGPT was genuinely useful here because it helped me generate multiple metaphor options quickly, then narrow to the one that was most likely to work in an image generator.
When it was time to write prompts, the best ones were structured, not clever. They had three parts: style rules I reused every time, composition rules like square versus landscape and where I needed negative space for text, and then the specific objects and their relationship. A typical prompt included constraints like keep it minimal, make the subject fill most of the frame, avoid tiny details, increase contrast, avoid faces, no extra objects. The goal wasn’t artistry. The goal was clarity.
Finally, I treated the first image like a draft. I’d generate it, then do a quick critique. Does it read instantly? What is confusing? What is too subtle? Is it going to wash out on the projector? Then I’d revise one thing and regenerate. Sometimes the tool followed revisions perfectly. Sometimes it didn’t. In those cases it was often faster to start fresh with a new prompt that baked in the fix instead of trying to “edit” my way out.
Best practices that emerged
Keeping scenes simple matters. If your image has three main objects, you’re usually fine. If it has eight, you’re likely in trouble. The generator will happily add details you didn’t ask for, which is great for art and terrible for slides.
Avoid text inside generated images when you can. Even when text “works,” it can come out slightly off. Your slide already has text. Let your slide do the reading and let the image do the meaning.
When you need people, hands work better than faces. Faces can shift tone fast, and they can also get weird. Hands show action, control, collaboration, and choice, which is exactly what you often want in a talk.
Use bold symbols, not subtle symbolism. A lightbulb beats “a thoughtful expression.” A checkmark beats “a feeling of success.” A warning triangle beats “a vaguely ominous mood.” Projector slides are not the place for subtle.
Do the squint test. If you squint and you can still tell what it is, you’re in good shape. If it becomes a blob, you need fewer elements and stronger contrast.
Know when to pivot. Some concepts will fight you. If you try a few times and the generator keeps missing the key idea, it’s often better to change metaphors than keep wrestling. That’s not failure, that’s design.
What I’ll do differently next time
Next time I’m going to build a small prompt library for recurring slide types, section transitions, benefits, risks, and proposals. Once you have a handful of prompts you trust, you can produce new decks faster and keep the visuals consistent without reinventing the wheel.
I’ll also start earlier. This workflow invites a little iteration, but the payoff is that you end up with visuals that match your message, instead of visuals you found after a long search.
