higgsfield-creative

这段内容是审核时保存的快照。外部来源才是完整且最新的版本。

---
name: higgsfield-creative
description: Generate AI ad video and image creative via the Higgsfield MCP server (project-scoped, configured in .mcp.json). Wraps Higgsfield's image and video models with ad-shaped prompt templates (9:16 reel, 1:1 square, 4:5 portrait) and routes to the right product (Soul, Speak, Ace, etc.) based on intent. Use when the user wants AI-generated b-roll, motion shots, or character-led ad video without filming.
metadata:
  tags: ads, video, ai-video, higgsfield, creative, generate
---

## When to use

- Generating cinematic / mood / b-roll ad video for PFC
- Generating premium product / lifestyle imagery (alternative path to nanobanana when we want Higgsfield's specific models)
- Producing character-led talking shots without filming a real person (compare HeyGen)
- Animating a still product shot into a 5-10s motion clip

## Prerequisites

1. The project's `.mcp.json` must include the higgsfield server (it does):
   ```json
   { "mcpServers": { "higgsfield": { "type": "http", "url": "https://mcp.higgsfield.ai/mcp" } } }
   ```
2. The OAuth handshake must be completed (run `/mcp`, select higgsfield, sign in).
3. Confirm with `/mcp` that `higgsfield: connected` before invoking this skill.

## Discovery step (mandatory)

Higgsfield's MCP tool surface changes as their products evolve. Before any
generation call, run a tool-discovery pass:

```
ToolSearch query="higgsfield" max_results=30
```

Then read each surfaced tool's input schema to know:
- What model/product it routes to (Soul ID, Ace, Speak, Veo3, Kling, Sora2, etc.)
- What aspect ratios + durations it supports
- Whether it accepts a reference image (for product shots / character-locked shots)

If the only tools available are `mcp__higgsfield__authenticate` and
`mcp__higgsfield__complete_authentication`, the OAuth handshake hasn't finished —
stop and tell the user to complete it.

## Ad-shaped prompt templates

PFC creative dimensions (Meta ad placements):

| Placement | Ratio | Resolution | Duration |
|---|---|---|---|
| Reels / Stories | 9:16 | 1080×1920 | 6–15s |
| Feed video | 4:5 | 1080×1350 | 6–15s |
| Square feed | 1:1 | 1080×1080 | 6–15s |

### Prompt formula (6-component)

```
[CONTEXT]    – the moment / use case (post-workout meal log, 5am kitchen scan)
[SUBJECT]    – who or what (real-body adult, hand holding phone, bowl from above)
[SETTING]    – environment + light (warm-window kitchen, soft directional)
[PROPS]      – explicit prop list pulled from the persona's scene→prop table
[STYLE]      – aesthetic anchor (editorial wellness, Bevel-adjacent, muted)
[TECHNICAL]  – framing, lens, motion (35mm, slow push-in, shallow DOF, no titles)
[REALISM]    – `ugc` (default for lifestyle) or `editorial` (default for studio/beauty/world plates)
```

`[PROPS]` is non-negotiable for any persona-driven shot. Pull it from the
scene→prop coherence table in the persona doc (e.g.
`ad-assets/avatars/Maya/persona.md`). If the scene doesn't appear in the
table, **stop and ask** — don't guess props.

### [REALISM] presets

**`ugc`** — default for lifestyle / kitchen / studio-exit / desk / couch / city shots:

```
shot on iPhone 15 Pro, default camera app, 1x lens
natural sensor noise, no color grade, slight overexposure on window highlights
single light source — window or one overhead pendant, NOT studio fill
composition slightly off-center, head not centered, occasional crop at top of frame
visible imperfections — dust on counter, water ring under glass, used tea bag
subject mid-action, not posed; not always making eye contact with lens
phone-arm sometimes visible in mirror / window reflection
skin texture preserved — pores, faint lines, no smoothing, no whitening
```

**`editorial`** — default for studio block (front-neutral / 3q / chin-up / beauty closeup) and world plates:

```
35mm full-frame look, shallow DOF, calm directional natural light
single key + bounced fill, no harsh shadows
subject calmly aware of lens, slight smile or candid laugh
clean