# Wan 2.2 Prompting Playbook for AutoShortsBot

Purpose: local Wan 2.2 video generation for a semi-automatic Retention Studio, especially mechanism-reveal Shorts around Hidden Money & AI Systems.

## Baseline findings

Sources checked:

- ComfyUI official Wan2.2 workflow docs: native Wan2.2 supports T2V and I2V templates, with the 5B TI2V workflow using `wan2.2_ti2v_5B_fp16.safetensors`, `umt5_xxl_fp8_e4m3fn_scaled.safetensors`, and `wan2.2_vae.safetensors`.
- Wan2.2 model docs/model card: 5B TI2V supports text-to-video and image-to-video at 720p/24fps on consumer GPUs such as RTX 4090; practical local A5000 16GB testing should still ramp below 720p first.
- Wan2.1 reference examples: negative prompt includes static/blur/subtitles/low-quality/deformed/body errors and uses guidance around 5.0 with UniPC-style sampling.
- Local tests on RTX A5000 16GB: native Wan2.2 + scaled T5 works; wrapper rejects scaled T5. T2V is weak for concrete subject control; I2V is materially better but needs a strong keyframe.

## Strategic rule

For production candidates, prefer:

```text
Curated/AI-generated keyframe -> Wan2.2 Image-to-Video -> deterministic edit/captions/SFX
```

Avoid relying on:

```text
Text prompt -> full final video
```

T2V is acceptable only for rough exploration, abstract atmosphere, or non-critical B-roll.

## Prompt structure

Use one compact paragraph, ordered from most important to least important:

1. Subject identity and count
2. Main action/motion
3. Environment
4. Camera behavior
5. Lighting/exposure
6. Style/aesthetic
7. Quality/control constraints

Template:

```text
[shot type] of [single subject] [doing one specific motion] in [specific environment].
Camera [one motion only]. [Lighting/exposure]. [Style/texture]. Preserve [important identity/details].
No text, no logos, no watermarks.
```

## Good I2V prompt pattern

```text
Photorealistic vertical wildlife video of one cute brown rabbit hopping forward through a sunny green meadow. The rabbit makes small natural hops, ears bouncing gently, fur remains realistic and consistent. Camera tracks smoothly from the side at low angle. Bright daylight, balanced exposure, shallow depth of field, natural colors, high quality wildlife footage. Preserve the same rabbit from the starting image. No text, no logo, no watermark.
```

Why this works:

- one subject
- one motion
- one camera move
- explicit realism/style
- explicit identity preservation
- no baked text

## Bad prompt pattern

```text
A beautiful cinematic scene with a rabbit jumping around a magical meadow, dynamic camera, epic lighting, cute, high quality, realistic, detailed, amazing.
```

Problems:

- vague motion
- no direction/path
- contradictory style risk
- no identity preservation
- no exposure control
- no artifact controls

## Negative prompt baseline

Use a reusable negative prompt; adapt per subject:

```text
cartoon, animation, deformed subject, extra limbs, extra ears, duplicated subject, disappearing subject, melting body, distorted face, bad anatomy, unrealistic movement, flickering colors, abstract, blurry, low detail, static image, camera shake, overexposed, underexposed, dark frame, text, subtitles, captions, watermark, logo, unreadable letters, fake UI text
```

For mechanism/finance videos add:

```text
fake numbers, fake labels, readable UI text, distorted charts, hallucinated interface text, brand logos
```

## Motion guidance

Wan obeys simple motion better than complex choreography.

Prefer:

- `slow push-in`
- `smooth side tracking shot`
- `gentle parallax move`
- `small natural hops forward`
- `object rotates slowly 20 degrees`
- `layers peel away one by one`

Avoid:

- multiple simultaneous camera moves
- long multi-step actions in one clip
- exact physics instructions
- fast cuts inside the prompt
- complex object transformations

## Image-to-video keyframe requirements

A good keyframe should have:

- full subject visible if movement matters
- side or three-quarter pose for directional motion
- enough empty space in the movement direction
- no readable text, no license plates, no UI labels
- clean background and strong silhouette
- same aspect ratio as target video where possible
- subject not cropped at important limbs/edges

For a hopping rabbit, use a side-view mid-hop or ready-to-hop pose, not a frontal sitting rabbit.

## Settings ladder for local A5000 16GB

Quality has priority, but avoid wasting hours on bad prompts:

1. Keyframe validation image/video still
2. I2V probe: `416x736`, 49 frames, 10 fps, 12-16 steps
3. Better candidate: `480x832`, 49 frames, 10-16 steps
4. Higher candidate only if subject/control is good: `576x1024`, 49 frames, 12-20 steps

Use native Wan2.2 with:

```text
model: wan2.2_ti2v_5B_fp16.safetensors
text encoder: umt5_xxl_fp8_e4m3fn_scaled.safetensors
vae: wan2.2_vae.safetensors
sampler: UniPC/uni_pc where available
cfg: around 4.5-6.0 as starting range
```

## Style direction for Hidden Money & AI Systems

Recommended channel look:

```text
photorealistic macro investigation + clean mechanism overlays
```

AI-video clips should provide real-world texture, not explain everything alone.

Visual palette:

- dark graphite / black UI base
- warm amber warning highlights
- electric cyan system/data lines
- red only for risk/reveal moments
- white kinetic captions

Clip style:

- realistic hands/devices/cars/desks/shops/offices
- clean Swiss/European environments when relevant
- no fake readable screens
- no generated text in video assets
- text and numbers added later by deterministic renderer

Mechanism animation style:

- x-ray layers
- glowing system nodes
- arrows/flows
- price/risk/time layers
- subtle 3D parallax

## Production rule

Generate AI clips without readable text. Add captions, labels, numbers, UI and arrows later with the deterministic renderer.