Video ad prompts
AI video models are good at light, texture, motion and atmosphere. They are bad at exact logos, readable text, hands holding small objects, counting, and keeping a product identical from shot to shot. Write prompts that use the strengths and route around the weaknesses. The ad is assembled in an editor, the model only supplies the footage.
Hard rules
- The model never draws the logo or the on-screen text. Logos, prices, claims, captions and the end card are added in the edit from the real files. Prompts say "no text, no logo, no letters" unless a model is being tested on purpose.
- The product comes from a reference image whenever the model allows it. Use image-to-video or a start frame made from a real product photo. A product described in words alone will drift in shape, color and label.
- Nothing invented in the ad. No fake reviews, results, awards, prices, before and after claims or medical effects. A testimonial prompt produces a scene, the words come from a real customer with permission, or the clip is labeled as a dramatization.
- No real people's likeness, no real brands, no copyrighted characters or places presented as endorsements. AI actors are generic and the ad must respect each platform's rules on synthetic or altered media. Tell the user to use the platform's AI label where required.
- One shot, one prompt, one main movement. Do not chain three actions in one generation. Two movements at once confuse the model and the viewer.
- Readable without sound. Every shot must make sense muted. Audio in the prompt is a bonus, never the carrier of the message.
- Model features change with every release. Check the current limits (clip length, audio, reference images, start and end frames, aspect ratios) before promising a result. See
references/model-notes.md.
Inputs to collect
- The product or service, and 1 to 3 real reference photos (packshot, in use, texture close-up).
- The audience and the one message the ad must leave (one sentence).
- The use: product ad, brand film, UGC, testimonial or launch. If a storyboard exists (from
ad-script-storyboard), use it shot by shot. - Target model(s) and plan limits, aspect ratio (9:16, 1:1, 4:5, 16:9), total length.
- Brand look: palette in plain words (warm cream, deep green), materials, mood, what the brand never looks like.
- Constraints: setting, season, people or no people, ages and diversity wishes, anything the legal team bans.
Workflow
Step 1. Fix the message and the shot list
Write the message sentence: "For [audience], [product] [benefit] because [real fact]." If no storyboard exists, list 4 to 7 shots with their role (hook, product, proof, use, end) and duration. Each shot should be 1 to 4 seconds in the final cut, even if the model generates 5 to 10 seconds. You will trim.
Step 2. Choose the route per shot
- Text-to-video: atmosphere, places, generic people, abstract textures, opening hooks.
- Image-to-video from a real photo: anything where the product must be right. The prompt then describes only motion, camera and light, not the product's shape.
- Start and end frame: reveals and transformations (closed box to open box, empty table to set table).
- Shoot it for real: screen recordings, exact hands-on demos, the founder speaking. Say so. Not every shot should be generated.
Step 3. Write each prompt with the anatomy
Use this order, in plain sentences, 40 to 120 words:
- Shot type and lens feel: macro, close-up, medium, wide, top-down, handheld phone.
- Subject: who or what, specific and physical (material, color, age range, clothing).
- Action: one verb, one movement, with its speed.
- Setting: place, time of day, a few concrete props.
- Camera movement: static, slow push-in, orbit, dolly, tracking, handheld sway. One only.
- Light: source and quality (soft window light from the left, low golden sun, hard flash).
- Style: commercial, documentary, phone footage, film grain, color grade words.
- Audio (if the model supports it): ambient sound, one effect, or spoken line in quotes, short.
- Exclusions: no text, no logos, no watermark, no extra fingers, no distorted product.
- Technical: aspect ratio, duration, frame rate if available.
Start from the closest prompt in references/prompt-library.md and adapt. Never paste a library prompt unchanged.
Step 4. Adapt per model
Rewrite for the target model using references/model-notes.md: some models follow long cinematic descriptions, others work better with short, literal sentences and a strong reference image. Keep one master prompt and note the model-specific version below it.
Step 5. Plan generation and selection
For each shot: number of generations (3 to 4 to start), seed or variation notes, what counts as a pass. A take passes when the product matches the reference, motion has no warping, hands and faces hold, nothing reads as text, and the action lands in the first second of the usable part. Keep a simple log: shot, prompt version, take, pass or fail, reason.
Step 6. Hand over to the edit
List what the editor adds: logo, captions, claims with their source, price if any, music, sound effects on each visible action, end card held at least 1.5 seconds. Point to ad-formats-delivery for safe zones and export.
Output template
# Video prompts: [Product], [use], [ratio], [length]
Message: [one sentence]
Target models: [..] | References supplied: [list of photos]
## Shot list
| # | Role | Final length | Route | Model |
## Shot [#]: [role]
Master prompt:
> [40 to 120 words, anatomy order]
[Model] version:
> [adapted]
Reference: [photo file] | Takes: [n] | Pass criteria: [..]
## Added in the edit
Logo, captions, claims (with source), music and sound, end card.
## Generation log
| Shot | Prompt version | Take | Pass | Reason |
Quality checklist
- Every prompt has one main action and one camera movement.
- No prompt asks the model to write text or draw the logo.
- Product shots use a real reference image or are marked as approximate.
- No invented claim, review, result or price anywhere in the plan.
- No real person's likeness, no real brand, no protected character.
- Each shot reads with the sound off.
- Aspect ratio and duration are stated in every prompt.
- Model-specific versions respect the model's current limits, checked by the user.
- The edit list covers logo, captions, sources, sound and end card.
Files
references/prompt-library.md: 40 prompts by use (product ad, brand film, UGC, testimonial, launch).references/model-notes.md: how to adapt a master prompt to Veo, Kling, Sora and Seedance, and what to check before each project.examples/example-prompt-set.md: a full prompt set for a fictional ceramics brand.