Prompt Patterns That Survive the Pro to Mini Jump

When a model family ships a Mini or Fast variant, half your prompts stop working and half keep landing. The half that keeps landing follows four patterns. This post walks through each one with a before and after you can run on bytedance/seedance-2.0/text-to-video today, since Mini is still in preview. The phrasing that survives on 2.0 is the phrasing most likely to transfer to Mini at launch.

Pattern one: camera direction wording

Mini variants tend to lose some narrative coherence but keep camera reasoning intact. Direct camera verbs survive the shrink. Vague cinematic adjectives do not.

Before: a cinematic aerial shot with epic movement over a coastal highway

After: aerial shot pushing forward over a coastal highway at sunset, golden hour side light, dolly right slow

The second version names the camera move (pushing forward, dolly right slow), the height (aerial), and the lighting direction. Each phrase maps to a concrete operation the model can plan. Cinematic and epic map to nothing the smaller variant can reliably reconstruct.

Pattern two: subject-first ordering

Smaller models budget attention more aggressively. Whatever you put in the first 12 words gets the most weight. Lead with the subject and the action, not the setting and the mood.

Before: In a dusty workshop bathed in warm afternoon light filtering through high windows, a blacksmith hammers a glowing horseshoe on an anvil

After: A blacksmith hammers a glowing horseshoe on an anvil, dusty workshop, warm afternoon light through high windows

The subject (blacksmith), the verb (hammers), and the object (horseshoe on an anvil) hit first. The setting follows as modifiers. On a Mini variant this ordering protects subject identity across the full duration of the clip.

## Pattern three: multi-shot delimiters

If you want a clip to read as two or three beats, do not write one long sentence. Use explicit shot markers. Mini variants handle delimited shots better than implied transitions.

Before: a chef plates a dish then the camera moves to a customer eating it then they smile

After: Shot 1: a chef plates a dish, close on the plate. Shot 2: a customer takes the first bite, mid two-shot. Shot 3: the customer smiles, tight on the face.

The delimiters give the model permission to cut. Without them, smaller variants tend to morph between beats inside one continuous frame, which looks broken.

Pattern four: audio cues that match the visual

Seedance ships native audio default on. Mini will most likely keep that behavior. Audio prompts that name a single concrete sound transfer cleanly. Adjectival audio prompts do not.

Before: cinematic music with epic feel

After: low cello drone, single footstep echo, wind through trees

The second version names three distinct audio elements. Each is something the model can produce. Cinematic and epic are not audio descriptions; they are review words.

A working test

Run the same prompt on 2.0 today. If it lands, save it. When Mini ships, run the same prompt and compare. Prompts that survive the shrink are the ones you build a campaign around.

01import { fal } from "@fal-ai/client";
02
03// Seedance 2.0 stand-in while Mini is in preview
04const result = await fal.subscribe("bytedance/seedance-2.0/text-to-video", {
05  input: {
06    prompt: "A blacksmith hammers a glowing horseshoe on an anvil, dusty workshop, warm afternoon light through high windows, camera pushes in slow.",
07    resolution: "720p",
08    duration: 6,
09    aspect_ratio: "16:9",
10    generate_audio: true
11  },
12  logs: true
13});
14
15console.log(result.data.video.url);

What to cut from your prompt library before Mini lands

Three categories of phrasing tend to break on the shrink.

Long adjective chains that describe a vibe. Moody, atmospheric, dreamlike, ethereal lighting reads as one undifferentiated cluster to a smaller model.

Implicit time of day. Late afternoon, just before the golden hour, when shadows start to stretch is three competing time markers. Pick one.

Metaphorical action verbs. The light dances on the water underspecifies. Sunlight ripples on the water surface gives the model a specific operation.

Clean these out of your prompt bank now. The result is shorter prompts that land more often on 2.0 and that should carry over to Mini once it ships. We will run a side by side on the same prompt set the day Mini opens up.