Most bad AI video ads aren't a model problem. They're a prompt problem. A good prompt is a shot list, not a wish, and once you write them that way, the same tools that gave you warped faces and stock-footage vibes start giving you ads you'd actually run.
Why your video ad prompts fall flat
The usual mistake is treating the model like a search engine: "a video ad for my coffee brand, cinematic, high quality." You'll get something. It'll look like everyone else's something. Vague in, generic out.
Video models reward specificity in a very particular order. They care most about subject, action, camera, and setting — roughly in that order. Adjectives like "stunning" and "professional" do almost nothing. Concrete nouns and verbs do the work. "A hand pours espresso into a glass of ice, the crema swirling as it hits" beats "beautiful cinematic coffee shot" every single time.
The 6-slot recipe that works across models
Whether you're writing a veo prompt, a kling prompt, or a seedance prompt, the same skeleton holds. Fill these six slots in order and you've written 90% of a good prompt:
- Subject — who or what, described concretely (age, material, color, one distinguishing detail).
- Action — one clear motion. One. Not three things happening at once.
- Camera — shot size and movement: "slow push-in," "handheld tracking," "locked-off macro."
- Setting & light — where, and what the light is doing ("soft window light from the left," "hard noon sun").
- Mood — two or three words, not a paragraph ("calm, warm, unhurried").
- Style / format — "shot on 35mm," "iPhone vertical, natural," "editorial product film."
Here's the skeleton filled in for a real product spot:
Macro shot of a matte black skincare bottle on wet slate. A single droplet rolls down the label and off the base. Camera slowly pushes in, locked and steady. Soft diffused light from the upper left, dark background. Mood: clean, premium, quiet. Shot on a macro lens, shallow depth of field, subtle film grain.
Notice there's no "amazing" or "4K masterpiece" anywhere. Every word is a decision the model can act on.
What each model actually wants
The 6-slot recipe is your base. Each engine has a personality on top of it, and matching your ai video prompting to the model is where quality jumps.
Veo
Veo is the most "director-literate" of the bunch and it handles native audio, so your veo prompt can name the sound. Describe ambient audio explicitly ("faint cafe murmur, the hiss of the machine") and it'll build a soundscape instead of leaving you a silent clip to score later. Veo also respects real cinematography vocabulary — dolly, rack focus, depth of field — so spend those words freely. It's your pick when you want a shot that feels filmed rather than generated.
Kling
Kling is strong on human motion and physical realism, which is exactly where cheap AI video usually falls apart. A good kling prompt leans into believable body movement and keeps the action count low — one gesture, fully described, holds together far better than a busy scene. It rewards a clear beginning and end state: "she picks up the mug, brings it to her lips, sets it down." Give it a journey, not a freeze-frame.
Seedance
Seedance is fast and punchy, great for snappy social cuts and energetic product motion. A strong seedance prompt is a little more compact — it likes momentum, so front-load the action and don't over-qualify. Use it when you need several quick variations to test, not one precious hero shot.
The anti-slop checklist
"AI slop" is a look, and viewers now clock it in under a second. Most of it comes from things you can prevent in the prompt. Before you hit generate, scan for these:
- Kill baked-in subtitles. Never write "with text on screen" or "captions." Models bake warped, misspelled text right into the frame. Add your text in edit, on top of a clean plate.
- Limit faces and hands in motion. If a human isn't essential, shoot hands-only or over-the-shoulder. When you do need a face, keep the head movement small and the shot short.
- One action per clip. Every extra simultaneous motion is another thing the model can render wrong. Chain simple clips instead of asking for one complex one.
- Anchor the product. For anything with a label, describe it precisely and keep the camera close enough that the model can't reinvent it. Loose framing is where your logo turns into gibberish.
- Name the light. "Cinematic" is a light description the model ignores. "Hard side light, deep shadows" is one it obeys.
Iterate like an editor, not a gambler
Don't reroll the whole prompt when one thing is off. Change a single variable and regenerate. Too static? Adjust only the camera line to "slow push-in." Wrong energy? Touch only the mood words. Treating it like a slot machine burns budget and teaches you nothing; treating it like a mixing board teaches you exactly which lever moves which result.
Keep your winners. When a prompt lands, save the whole thing verbatim as a template and swap only the subject next time. A prompt that nailed one product shot will usually nail the next, and you build a personal library that gets better every week.
Let the tool write the boring parts
You don't have to hand-tune every clause forever. This is exactly what Bloopo is for: it connects to Claude or ChatGPT as a tool, takes a plain sentence like "a calm 20-second ad for my skincare bottle," and expands it into the model-specific prompts above — routing to Veo, Kling, or Seedance as needed, stripping out the slop triggers, keeping your product label accurate, and showing you the price (a ~20-second ad runs around $4.87) before it spends a cent.
Point your assistant at https://mcp.bloopo.ai/mcp, describe the ad in one line, and let it handle the prompt craft while you judge the results. That's the fastest way to turn these recipes into finished video without becoming a full-time prompt engineer.