One Brief Three Forks Sora Veo Gpt Image Youtube Thumbnail
You can brief Sora, Veo, GPT Image, and a YouTube thumbnail with the same product facts and still get four unrelated objects. That is usually a brief problem, not a model problem.
Write the facts once. Fork the control surface. Video models want shots and motion. Image models want layout and readable labels. Thumbnails want a face-sized subject and five words, not a paragraph. Matching fork prompts are landing in the PromptDig library (Browse more prompts). Keep the trio next to each other so you do not rewrite the product from memory.
This is the same skill as prompting Midjourney and Flux from one still. Different dialects. One brief.
Write a photographer's brief, not a slogan
Start with facts a stranger could shoot:
- Product: exact object, color, what must stay readable (logo, label, UI)
- Job of the piece: ad, explainer, thumbnail
- Audience in one line
- Setting and time of day
- Camera: distance, angle, what is in frame
- Motion (for video): what moves, what must not
- Must include / must not include
- Words that may appear on screen (exact strings only)
That document is model-agnostic. If you cannot fill it, you are not ready to generate. You are still brainstorming.
A skeleton for the shared brief:
Product: [object, color, logo rules].
Job: [15s product ad / infographic / thumbnail].
Audience: [who].
Setting: [place, time, surfaces].
Camera: [distance, angle, lens feel].
On-screen text allowed: [exact strings, or NONE].
Must not: extra products, fake UI, unreadable labels, celebrity lookalikes.
Lock: product silhouette and brand colors.
Everything after this is translation.
Fork A: Sora and Veo, as shot lists
Sora and Veo want time. If you paste a still-image paragraph, you will get a looping atmosphere with a product that drifts. Write shots.
Keep the ad short in the prompt even if the tool allows longer. Three to five shots is plenty for a product beat:
- Wide: product in a real place
- Insert: the one detail that proves the claim (texture, latch, screen state)
- Hands or use: one action
- Lockup: product plus the allowed on-screen line
Say what is still. "The bottle does not morph. The label does not reprint. No extra bottles appear." Video models love to spawn twins.
Put motion in verbs attached to objects: "steam rises," "hand slides the cap," "camera pushes in." Do not say "cinematic" and hope. Specify duration per shot if the tool accepts it. If it does not, specify the action that implies duration.
Sora and Veo will still differ. Veo often follows a shot list more literally. Sora often adds style you did not ask for. Counter that in the fork: for Sora, add "documentary, no slow-motion hero hair, no extra props." For Veo, add the spatial language (left, right, foreground) the way you would for a literal image model.
Audio is a separate brief. If you need a voiceover, write the VO as exact words and say "no additional claims." If you do not need audio, say silence or ambient only. Models will invent a slogan if you leave the soundtrack unspecified.
Fork B: GPT Image, as an infographic spec
GPT Image (and similar still models that can letter) wants a layout, not a camera move. Turn the same brief into a poster spec:
- Canvas: 1:1 or 4:5, not 16:9 unless it is a slide
- Regions: title band, three callouts, product hero, footer
- Exact strings for every label
- "Do not invent extra statistics. If a number is not in the brief, omit the callout."
Infographics fail when the model is asked to "show the benefits" and it fabricates percentages. Put the only allowed numbers in the brief. If you have none, use qualitative callouts: "washes in cold water," not "30% more."
Ask for real letterforms and then proof them. If a label is legal or brand-critical, composite the type in a layout tool and use the model for the hero object only. The Midjourney-vs-Flux lesson still applies: some models are prettier, some are more literal. GPT Image is the literal fork in this stack.
Fork C: YouTube thumbnails, as a five-word poster
A thumbnail is not a smaller infographic. It is a billboard at 160 pixels wide.
Rules that belong in the fork, not in a style adjective list:
- One subject, large, cropped tighter than the ad
- Face or product, not both fighting
- Three to five words, max. Exact string from the brief
- High contrast between type and background
- No screenshots of fake UI, no tiny charts, no crowded storyboard
Reuse the product lock from the ad: same silhouette, same color, same angle if you can. Series thumbnails should look like a family. That is a lock list, the same way before-and-after image prompts lock identity.
If you need a person, use a model release you have, or illustrate. Do not ask for a lookalike.
Keep the forks in one folder
Name files by job, not by model: brief.md, fork-video.txt, fork-infographic.txt, fork-thumb.txt. When the product color changes, change the brief and regenerate the forks. Do not edit three prompts by hand and miss the fourth.
When a set actually ships, publish the brief plus the three forks (Share a prompt) so the next campaign does not start from a slogan. Browse the library on PromptDig (Browse more prompts) as the matching stacks land.
You do not need a new brand for each tool. You need one brief, then a shot list, a layout spec, and a thumbnail crop. Write the facts once. Translate the control surface.