OutcomeWrite image prompts that describe a photograph instead of a concept.
Text prompting and image prompting are different skills. Text models want instructions. Image models want description: they respond to what a scene looks like, not what it means.
"An image representing calm and focus" produces stock-photo mush. What works is describing an actual photograph: subject, setting, what they're doing, time of day, light source and quality, camera position and distance, lens character, color palette, mood. "A woman in her forties sitting on the edge of an unmade bed at 5am, lit only by a phone screen, shot from across the room on a 35mm lens, muted blues, shallow depth of field." That's a picture. The first one was a mood board.
Order matters. Most models weight the front of the prompt more heavily, so lead with the subject. And negatives are as useful as positives: name what you don't want in frame.
One firm limit: don't generate images of real, identifiable people, and don't generate anything that reads as documentary evidence of something that didn't happen. That's not a stylistic guideline.
Drill20 min
Take one image you need. Write the concept version and the photograph version. Generate both in an image-capable AI program. The gap teaches the lesson faster than any amount of reading about it.
Prompt: Turn a concept into an image prompt
Turn my concept into a proper image prompt.
CONCEPT: [what you want to convey]
USE: [where it goes: social post, header, thumbnail]
FEELING: [the emotional register]
Write it as a description of an actual photograph, in this order:
1. Subject: who or what, specifically, and what they're doing
2. Setting and time of day
3. Light: source, direction, quality
4. Camera: distance, angle, lens character, depth of field
5. Color palette
6. Mood
Then give me a NEGATIVE list: what must not appear.
Do not use abstract words like "empowering," "modern," or "dynamic." They mean nothing to an image model. Describe what a camera would see.
Paste into ChatGPT, Claude or Gemini. Replace every [bracketed] placeholder with your own details first, then review the edited prompt before you send it.
Visual Consistency
25 min
OutcomeMake ten images look like they came from one brand.
One good image is luck. Ten images that look related is a brand. The difference is a style block: a fixed paragraph you append to every image prompt, specifying palette, lighting, camera treatment and mood.
Build it once. Three or four sentences: your color range, your light (hard or soft, warm or cool, natural or artificial), your typical framing and lens feel, your level of realism. Then every prompt becomes [subject description] + [style block], and your feed starts looking like it belongs to one person.
Practical constraints worth deciding early: whether people appear in your imagery at all, whether you'll use text-in-image (models still mangle it), and what your images will never look like. The aesthetic you're deliberately avoiding is as defining as the one you're chasing.
Drill25 min
Generate four unrelated images with the same style block appended. Put them side by side. Adjust the block until they read as a set, then save it next to your voice profile.
Prompt: Build a visual style block
Help me build a reusable visual style block.
MY BRAND: [what you do, who for]
FEELING I WANT: [3-4 words]
WHAT I WANT TO AVOID: [the look you're deliberately not doing]
Ask me 5 questions to pin down palette, light, framing, and realism level. One at a time.
Then produce a STYLE BLOCK of 3-4 sentences I can append to every image prompt, plus a NEGATIVE line of things that should never appear.
Finish by writing three sample prompts (three totally different subjects) with the block appended, so I can test whether they come out looking related.
Paste into ChatGPT, Claude or Gemini. Replace every [bracketed] placeholder with your own details first, then review the edited prompt before you send it.
Decks and One-Pagers
25 min
OutcomeGet a presentation to draft stage in fifteen minutes instead of three hours.
Tools like Gamma will turn an outline into a formatted deck in about a minute. That is genuinely useful and also genuinely dangerous, because a beautiful deck full of vague slides is worse than an ugly deck with a real argument. It takes longer for you to notice it's empty.
So write the argument first, in text, before you generate anything. One line per slide stating the claim that slide makes. If you can read your outline top to bottom and it makes a coherent case, the deck will work. If your outline reads "Introduction / About Us / Our Solution / Benefits / Next Steps," you don't have an argument, you have a shape.
Then generate, then cut. The generated version will have too many slides and too many words on each. The editing is where the value is, and it's the part AI can't do for you because it doesn't know what your audience already believes.
Drill25 min
Write a slide-by-slide claim outline for a presentation you actually need to give. Read only the claims aloud. If it does not hold together as an argument, fix that before you generate a single slide.
Prompt: Build the argument first
Before we design anything, build the argument.
PRESENTATION: [topic]
AUDIENCE: [who, and what they believe right now]
DECISION I WANT: [what they should do at the end]
TIME: [minutes]
Produce a slide-by-slide outline where each line is THE CLAIM THAT SLIDE MAKES: a full sentence that asserts something, not a topic label.
Rules:
- No slide called "Introduction," "About Us," "Overview," or "Benefits"
- Each claim must move the audience toward the decision
- Mark the single slide the whole talk hinges on
- One claim per slide, maximum
Then tell me which claims I'll need evidence for, and what kind.
Paste into ChatGPT, Claude or Gemini. Replace every [bracketed] placeholder with your own details first, then review the edited prompt before you send it.
AI Video and Voice: Where the Line Is
15 min
OutcomeUse synthetic media without damaging the trust you are building.
You can now clone your own voice, generate video of a person speaking, and produce an avatar that presents your content. The technology is good enough that the interesting question isn't "can I," it's "should I, and what do I owe the viewer."
A workable standard, stricter than the law in most places and safer for someone building a personal business:
Your own voice, your own words, disclosed: fine. Useful for narration, course audio, translating your content.
Your own likeness saying things you approved: fine, disclosed. Don't let a tool generate the words unsupervised.
Anyone else's voice or likeness: no. Not for a demo, not as a joke, not "just to test."
Synthetic people presented as real customers: never. This is the fabricated-testimonial problem with a face on it.
The strategic argument matters as much as the ethical one. Your entire business rests on people believing you're a real person telling the truth. Synthetic media spends that trust for a bit of production efficiency. Spend it carefully, and only where disclosure costs you nothing.
Drill15 min
Write your own three-line policy on synthetic media: what you will use, what you will not, how you disclose. Put it in your project instructions. Deciding in advance is much easier than deciding when a tool makes something tempting.
Prompt: Personal policy on AI-generated media
Help me write a personal policy on AI-generated media.
MY SITUATION: I run a personal-brand business where trust is the main asset. [add specifics]
Ask me where I stand on: cloned voice, AI video of myself, AI avatars, AI images of people, and disclosure.
Then write a 5-line policy I can put in my project instructions, plus the exact disclosure wording I'd use in each case.
Bias toward disclosure. Where a reasonable viewer would feel misled if they found out later, say so plainly.
Paste into ChatGPT, Claude or Gemini. Replace every [bracketed] placeholder with your own details first, then review the edited prompt before you send it.