通用提示原則（官方）

1. 官方 Skill：Prompt Engineering（原文）skills-prompt-eng

## Basics

Higgsfield models reward concrete, sensory prompts.

- **Subject + setting + style**: "a red fox curled in a snowy pine forest, golden hour, cinematic"
- **Camera**: lens (35mm, 85mm), angle (low, overhead), motion (dolly in, tracking shot)
- **Lighting**: rim light, neon glow, moody backlight
- **Style/medium**: oil painting, watercolor, photograph, anime, 3D render

Keep it under ~200 tokens. Models distort with very long prompts.

## Image-to-image

When passing `--image`, the prompt should describe what changes, not redescribe the input.

Bad: "a man with brown hair in a leather jacket holding coffee, made into anime"
Good: "transform into anime style, vibrant colors, soft cel shading"

## Image-to-video

`--start-image` anchors the first frame. Prompt describes motion.

- Verbs: zooms in, dollies left, sweeping pan, slow push, fast whip
- Subject motion: "the dancer spins", "smoke rises slowly"
- Don't redescribe the static frame — model already has it.

## Negative phrasing

Most models don't expose a `negative_prompt`. Phrase positively:
- Instead of "no blur" → "tack sharp"
- Instead of "no people" → "uninhabited landscape"

## Aspect ratio guidance

- `16:9` — landscape, cinematic
- `9:16` — vertical, social
- `1:1` — square, profile / icon
- `4:3`, `3:4`, `21:9` — model-dependent, check `higgsfield model get <jst>`

中文重點

2. 寫好影片提示的 6 步（官方網誌）blog:ai-video-prompt-mistakes

flowchart LR
  S1[1 場景與演員鎖定<br>年齡/體型/服裝/特徵] --> S2[2 先走位<br>位置、方向、相機側]
  S2 --> S3[3 相機=物理設定<br>視野/距離/運動類型]
  S3 --> S4[4 動作=分鏡清單<br>起點→終點、接觸點]
  S4 --> S5[5 物理與光線=真實事件<br>單一光源/方向/顏色]
  S5 --> S6[6 聲音與收尾<br>對白歸屬/音樂/膠片/排除項]

生成前檢查清單（原文）：

- **Character detail**: age, build, and clothing, backed by a reference or asset
- **Points of contact**: hands, fingers, and anything touching something else named explicitly
- **Emotion**: tied to a visible action, not a mood word
- **Light**: source, angle, and time of day named, or set in Cinema Studio
- **Movement**: an explicit turn, tilt, and angle, not an adjective
- **Shot purpose**: one job per shot, nothing in frame that doesn't serve it
- **Complex scenes**: storyboarded in Popcorn first

3. 防變形 10 招（官方網誌）blog:how-to-avoid-distortions-ai-videos

4. 提示遵從度（官方比較網誌）blog:ai-video-prompt-adherence-comparison

提示遵從度 = 模型輸出與提示要求的吻合程度。網誌以同一提示測試多款工具，結論與細節見來源；本庫不轉述未經核實的排名。

5. ChatGPT／MCP 通用範本（原文）blog:generate-ai-videos-chatgpt-2026

Generate a [length]-second [vertical or horizontal] video of [subject] [action] in [setting]. Camera: [shot size and movement]. Light: [direction and quality]. Mood: [feeling]. Before generating, show me the credit cost and wait for my confirmation.

本頁來源（存取日期 2026-10-08）