One dispatcher, many engines. Intents stay stable; engines are swappable. When an engine changes, this file changes — the prompts and pipelines that call it do not.
Stack prerequisites
This skill routes to engines you install separately. Missing engines degrade gracefully — the router picks the next-best available and says so.
| Engine | Role | Install |
|---|---|---|
| Higgsfield CLI + skills | Default image and video engine (paid account) | curl -fsSL https://raw.githubusercontent.com/higgsfield-ai/cli/main/install.sh | sh, then sync the vendor skills (higgsfield-generate, higgsfield-soul-id, higgsfield-product-photoshoot, higgsfield-marketplace-cards) |
| Gemini API key | Nano Banana image gen and edits (character/reference work) | GEMINI_API_KEY env var |
| Grok Imagine | Fast stylized images if you have a SuperGrok subscription | vendor CLI/app |
| HyperFrames | HTML video compositions, captions, TTS, audio-reactive motion | npx hyperframes + its skills |
| Descript MCP | Edit by editing text: cuts, filler removal, captions, publish | connect the Descript MCP server |
Step 1 — classify the request
Three axes decide everything downstream:
| Axis | Values |
|---|---|
| Deliverable | still image · single clip (≤15s) · assembled video (edited, multi-scene) |
| Identity consistency | none · product (same object every shot) · face (same person every shot) |
| Destination | social feed · marketplace listing · YouTube · website/hero · ad campaign |
If the user did not specify duration, aspect ratio, or destination, infer from context; ask only when the answer changes the engine choice.
Step 2 — dispatch
| Intent | Route | Why |
|---|---|---|
| Product/brand still | higgsfield-product-photoshoot | prompt templates for studio, lifestyle, hero, carousel |
| Marketplace listing set | higgsfield-marketplace-cards | compliant main image + secondary + A+ modules |
| Same face across shots | higgsfield-soul-id (train once) → higgsfield-generate with the Soul id | identity persists across images and video |
| Generic still, design, legible text | higgsfield-generate (GPT Image 2) | strongest text rendering |
| Character/reference image edit | Nano Banana via Gemini API | best reference-following for edits |
| Fast stylized still | Grok Imagine | included in a SuperGrok subscription, zero marginal cost |
| Single video clip, default | higgsfield-generate (Seedance 2.0) | best price/quality for social clips |
| Single video clip, cinematic | higgsfield-generate (Kling 3.0) | motion quality, camera control |
| Branded ad with avatar/product | Higgsfield Marketing Studio (via higgsfield-generate) | hooks, avatars, product placement in one flow |
| Captions, title cards, audio-reactive motion, scroll video | HyperFrames composition | deterministic HTML video, no generation cost |
| Cut, rearrange, remove filler, subtitle, publish | Descript MCP (import_media → prompt_project_agent → export) |
text-based editing beats timeline scrubbing |
Fallback order when an engine is missing: Higgsfield → Nano Banana (images only) → Grok Imagine (images only). If no video engine is installed, stop and say which install unblocks.
Step 3 — pipelines
30-second product ad
- higgsfield-product-photoshoot → 3 stills (hero, lifestyle, closeup)
- higgsfield-generate image-to-video on the hero (Seedance 2.0, 5s)
- HyperFrames: title card + price overlay + end card around the clip
- Descript: assemble, caption, export per destination aspect ratio
Faceless YouTube b-roll batch
- Script beats → one prompt per beat
- higgsfield-generate batch (Seedance 2.0, 5s each)
- Voiceover: HyperFrames TTS, or ElevenLabs if the user has it
- Descript: import all, sequence to narration, auto-captions
Talking-head with a consistent face
- higgsfield-soul-id: train once, store the reference id
- higgsfield-generate with the Soul id → presenter clips
- Descript: filler-word removal, captions, publish
Caption-heavy short (quote/hook format)
- Still from the dispatch table (or user-supplied)
- HyperFrames: audio-reactive captions, marker highlights, beat-synced motion
- Render via HyperFrames; no Descript needed
Quality gate before publishing
For anything going to a paid placement or a growth channel, run the Higgsfield virality
predictor (brain_activity) on the finished cut: hook strength, retention risk, distraction.
Score low → fix the first 2 seconds before touching anything else.
When you hit a ceiling
Credits exhausted, need a custom pipeline (LoRA, ControlNet, motion transfer), or per-clip
cost matters at volume → hand off to the video-engine-routing skill. It holds the decision
tree and current pricing for subscription vs pay-per-use vs cloud ComfyUI.