# ANOMALY Knowledge · The Video Prompting Bible

How to animate a still into a clip that looks real. Covers **Seedance 2.0** and **Kling 3.0**, plus the Higgsfield settings that decide whether your clip is usable. Motion is where most people's videos fall apart, so this matters as much as the image.

---

## The golden rule of AI motion

**Small, real motion beats big, impressive motion.** The moment a clip tries to do too much, hands warp, faces morph, and products melt. Every clip in an organic video is one simple beat. Animate one thing.

If you remember nothing else: **lock the camera, keep the action small, generate multiples.**

---

## Which model, when

| Model | Use it for | Notes |
|---|---|---|
| **Seedance 2.0** | Default for most clips. Build steps, reveals, product motion. | Takes a descriptive, timeline-style prompt. |
| **Kling 3.0** | A cheaper alternative, good for simple static-camera shots. | Different prompt style. **Hard limit: 2500 characters.** |

They take **different kinds of prompts.** Always tell your Claude which one you are using so it writes the right style. Never paste a Seedance prompt into Kling and expect the same result.

---

## The 8-part architecture (Seedance 2.0)

1. **Opening frame** exactly what is visible at 0:00
2. **Subject motion** how the person or product moves
3. **Camera motion** specific camera behavior (or explicitly none)
4. **Environment** setting, lighting, atmosphere
5. **Temporal breakdown** what happens across the seconds
6. **Audio design** ASMR elements, ambient sound
7. **Duration** target length
8. **Style** handheld UGC, cinematic, etc.

### Template
```
Opening frame: [exactly what the still shows].
Motion: [the one thing that moves, described plainly].
Camera: [locked / slow push in / slight handheld drift].
Environment: [lighting and atmosphere, matching the still].
[00:00-00:02] [what happens first]
[00:02-00:05] [what happens next]
Audio: [ambient or ASMR sound of the action].
Style: handheld iPhone UGC, natural light, realistic motion.
```

---

## Camera motion vocabulary

| Term | What it does |
|---|---|
| static / locked shot | No movement. Feels stable and premium. **The safest choice.** |
| slow push in | Draws the viewer toward the product or moment |
| handheld drift | Organic, real, authentic |
| rack focus | Blurs foreground then reveals the sharp product |
| orbital | Premium reveal showing all angles. Risky, drifts easily. |
| pull back reveal | Wide reveal of the full scene |

**For beginners: use "camera is locked" on most clips.** Moving cameras are where AI video breaks.

---

## Duration guide

| Length | Best for |
|---|---|
| 4-5s | One beat: a product reveal, a single action, an animated hook frame |
| 10s | A short sequence, two or three connected moments |
| 15s | A full arc, hook then build then reveal |

Most organic clips are built from **several 5-second beats cut together**, not one long generation. Short clips fail less and cut better.

---

## The proven pipeline

1. Generate the still with **Nano Banana Pro** (or GPT Image 2 for the character).
2. Pass that still as the **first frame** to Seedance or Kling.
3. Prompt the motion to start **from that exact frame**.
4. The model preserves your character and product and just adds movement.

Never generate video from text alone. Always animate a still you already approved. That is how you keep the character consistent across a whole video.

---

## Higgsfield settings that matter

- **Enhance: OFF.** Enhance rewrites your prompt and causes drift. This is the single most common mistake.
- **Aspect ratio: 9:16.**
- **Generate 2 to 3 per prompt** and pick the best. Sometimes more for a hard shot. Outputs vary significantly.
- **Kling prompts must stay under 2500 characters.**
- Keep prompts plain text. Structured JSON prompts fail on this endpoint.

---

## Motion rules that keep clips usable

1. **The action happens after the character finishes speaking, never during.** Mouth movement plus hand action at the same time is where faces break.
2. **Lock the camera unless you have a reason not to.**
3. **One action per clip.** Not "she cuts the fabric, then holds it up, then smiles." That is three clips.
4. **Hands are the weak point.** Keep hands doing one simple, slow motion. Avoid fast or complex hand work.
5. **Say what should NOT move.** "Background is static, camera does not move" prevents a lot of weirdness.
6. **Match the motion to the still's lighting.** Do not ask for a lighting change mid-clip.

### The reliable "making the product" prompt shape
```
The person is doing [the one action]. Camera is locked. Not talking.
Motion should be as smooth as possible. Background static.
```

---

## Sound

Organic videos are almost always cut to a trending song, so **generated audio usually gets replaced.** Still describe the ASMR of the action (the cutting, the pouring, the click) because it makes the model animate the action more convincingly, and because those natural sounds are sometimes worth keeping under the music.

---

## What to do when a clip comes out wrong

- **Face morphs:** the action was too big, or the character was speaking. Simplify the motion, lock the camera.
- **Hands warp:** slow the hand motion, or reframe so hands are partly out of frame.
- **Product changes shape:** your still was ambiguous. Regenerate the still with the product bigger and clearer, then animate again.
- **Scene brightens or shifts:** state "preserve the exact lighting from the first frame."
- **It just looks AI:** add handheld drift and slight imperfection, and cut the clip shorter. Short clips hide flaws.

**And the fix that always works:** ask your Claude, "Check the original prompt. Is there any reason it keeps doing [the problem]?" The cause is nearly always ambiguous wording in the prompt, not the generation.

---

## The confirm-first rule

Before anything generates, your Claude shows you every image prompt and every motion prompt in a numbered list, and waits. You scan them, fix what is off, then say go. Generation costs credits; reading a prompt is free. Never generate blind.