Project 1: The YouTube Thumbnail
Thumbnails live or die on three things: a face with a clear emotion, big readable text, and one bold focal object. AI handles the focal object and background. You handle the face (use your own photo) and the text (use a design tool, not the model).
Workflow:
- Open your reference: pull three top-performing thumbnails in your niche. Note the dominant color, the focal object, the emotion.
- Generate the background plate with a wide aspect ratio (16:9 or wider so you can crop).
- Composite your face in Photoshop, Photopea, or Canva. Add the text last.
Try this prompt for a "coding tutorial" thumbnail:
A glowing holographic Python logo exploding out of a laptop screen,
dark navy background, dramatic rim lighting, cinematic depth of field,
sharp focus on the logo, empty space on the left for text overlay,
16:9 aspect ratio, high contrast, vibrant cyan and orange accents
The "empty space on the left for text overlay" line is the secret. Models often ignore it, so generate four and pick the one with the cleanest negative space. Do not let the model render the text — it will misspell it, and you cannot pick the font.
Crop tight. Faces should occupy at least 30% of the frame. If your background looks soft after compositing, run it through a topaz upscaler or your tool's built-in upscale before exporting.
Project 2: The Pitch Deck Hero Image
Investors skim. A hero image on slide one buys you eight extra seconds of attention before someone opens email. The goal is not "art" — it is establishing what your company makes and the mood you want the room to feel.
Two patterns work:
- Product-in-context. Your product in a real environment, used by a real-looking person.
- Abstract metaphor. A visual that hints at the problem you solve (a tangled mess for "we untangle X").
Avoid the third pattern, which is "generic futuristic AI brain on circuit board." Investors have seen it 4,000 times this quarter.
Wide cinematic shot of a warehouse manager checking a tablet,
soft morning light through high windows, shelves of boxes in
soft focus behind her, muted blue and beige color palette,
shot on Sony A7 with 35mm lens, photojournalistic style,
shallow depth of field, copy space in the upper right
Aspect ratio: 16:9 for slides, 21:9 if you want letterboxing. Always generate at the highest resolution your tool offers — slides get projected, and pixelation reads as amateur. For the structure of the deck itself, the /courses/ai-pitch-decks-exec-presentations course covers narrative flow and which slides actually move the needle.
Project 3: The Ad Creative
Ads have a brutal job: stop the scroll, then make a promise. AI is excellent at the stop-the-scroll part. It is bad at the promise — that is your copy, your offer, and your brand.
The workflow that beats single-image generation: produce a set of five variants, all different visual hooks for the same product, and let the platform's algorithm pick the winner.
Five variants for a productivity app:
- Hero shot — clean product on a colored background.
- Lifestyle — a person using it in a relatable setting.
- Before/after — split-screen chaos versus calm.
- Pattern interrupt — the product floating in an absurd environment (in a soup bowl, on the moon).
- Testimonial frame — a face with negative space for a quote.
Run each prompt with the same color palette and lighting style so the set feels like a campaign, not five random images. Stitch the brand in post — logo, button, headline — using Figma or Canva. The AI gives you the canvas, never the call to action.
For writing the headline that goes on top, /courses/ai-writing-content-creation goes into hook-writing in more depth than belongs in a chapter on images.
Project 4: The Book Cover
Book covers are punishing because they have to work at thumbnail size on Amazon, at full size on a shelf, and the title has to be readable on a phone screen. The image is doing 60% of the selling. Get this wrong and your sales graph stays flat.
Three rules:
- Genre conventions are not your enemy. Romance has soft pastels and cursive titles. Thriller has high contrast and sans-serif. Learn the rules, then break exactly one.
- One subject. Covers with three or more focal points read as confused.
- Title space planned from the start. Generate with a deliberate dead zone.
A nonfiction business book example:
A single bonsai tree growing inside a clear glass cube,
clean white studio background, soft directional lighting
from the upper left, minimalist composition, the tree
positioned in the lower third with large empty space above,
2:3 portrait aspect ratio, sharp focus, editorial style,
muted natural greens and warm wood tones
The "positioned in the lower third with large empty space above" gives you room for the title. The "single subject on white" makes the thumbnail readable at 100 pixels tall on a phone. Use a separate typesetting tool — Affinity Publisher, InDesign, or even a Canva template — and never let the model render the author name.
Project 5: The Social Post
Social images are the lowest-stakes, highest-frequency project. They reward speed and a recognizable visual identity over polish.
Build a template, not a one-off:
- Pick three colors, one font, and one composition pattern. Write them down.
- Generate ten variations of background imagery using the same color palette and style descriptors.
- Drop each into your template (a square frame in Canva or Figma) with a quote, headline, or stat.
A reusable prompt skeleton:
[subject], in the style of [your locked-in art direction],
[your three colors] palette, [your locked-in lighting],
1:1 square aspect ratio, centered composition with
breathing room on all sides for text overlay, flat
illustration style, no text, no logos
You will produce fifty posts in the time it used to take to make five. The first instinct is to make every post a masterpiece. Resist it. Consistency of look beats peak quality on any single post — the feed is what people remember, not the individual image.
What to Take From All Five
Across every project, the pattern repeats: the model makes the canvas, you make the message. Plan the negative space, never let the model render critical text, generate in sets so you have choices, and finish in a real design tool. The chapter you just read is a template — copy the prompt structure, swap the variables, ship the work.