Higgsfield Motion Design Agent Skill

Higgsfield Motion Design

A full motion-design creation flow run through the Higgsfield MCP connector. Follow the steps in order, be concise and direct, and reply in the user's language. This skill is the guided *ad/brand-motion* pipeline — for the named camera/motion preset library (Explosion, Werewolf, Air Bending, etc.) use higgsfield-motion instead.

> Not a spec sheet. Model parameter enums (resolutions, modes, durations) come from the specs layer / models_explore — verify there (HARD RULE #3), don't hardcode them here.

QUICK FACTS

• Two flows: classicMD (smooth, elegant, cinematic) vs highMD (fast cuts, extreme dynamics, CGI energy) — pick before anything else →
• Ask all brief questions in ONE message — never split intake into rounds →
• Storyboard = one image: a single grid sheet with all N panels via GPT Image 2 — never N separate images →
• Final video = Seedance 2.0 (seedance_2_0); confirm the model id with models_explore if unsure →
• highMD rule: no realistic humans — silhouettes, chrome figures, or 3D abstract shapes only →
• highMD rule: the logo lock is a static hold proportional to clip length (~1s / ~2s / ~2–3s for 5 / 10 / 15s) →

STEP 0 — Determine the flow type

Identify which workflow applies before anything else:

• classicMD — standard ads, brand promos, service presentations, logo reveals, general atmospheric content. Smooth transitions, elegant typography, cinematic feel.
• highMD — sports promos, tech launches, music teasers, AI-capability demos, fashion drops. Extreme camera speed, aggressive cuts, peak dynamics; realistic people are replaced by silhouettes, chrome elements, or 3D abstract figures.

If the request makes the flow obvious, proceed silently. If ambiguous, ask once:

> "Which style fits your project better — Classic Motion (smooth, elegant, cinematic) or Hyper / Kinetic (fast cuts, extreme dynamics, CGI energy)?"

STEP 1 — Brief intake (one message)

Ask all of these in a single message (never split into rounds). Save every answer before proceeding:

• Existing assets? — Yes (upload logo / product photo / reference) · No (help me create the visual)
• Duration — 5s (teaser / logo sting) · 10s (standard post / stories) · 15s (promo / product video)
• Frame format — 16:9 (horizontal) · 9:16 (vertical Reels/TikTok/Stories) · 1:1 (square feed)
• Mood / style *(free input)* — e.g. energetic, minimalist, luxury, technological, atmospheric, aggressive, cinematic
• Brand / product name and tagline *(if any)*
• Storyboard frames — 6 (standard) · 8 (detailed) · 9 (maximum coverage)

STEP 2 — Asset handling

If the user HAS assets: when the client is an Apps UI-capable surface, call media_upload_widget immediately so they can attach the local file (remote MCP cannot read chat attachments). For a web media URL, call media_import_url first and pass the returned media_id. Then proceed to Step 3.

If the user has NO assets: generate a base visual with GPT Image 2 (generate_image, model gpt_image_2) — construct the prompt from brand name, mood, style, palette, aspect ratio. Display the result with job_display, ask "Does this work or want changes?", regenerate if needed, then proceed once approved.

STEP 3 — Generate the storyboard

The core creative step. Generate one storyboard sheet — a single image with all N panels in a grid (N = the count from Step 1: 6, 8, or 9). Do not generate N separate images.

Call generate_image once with GPT Image 2, passing the approved asset / base visual as a reference. Each panel must: stay visually consistent with the approved asset · represent a distinct moment (opening → build → climax → resolution → logo lock) · show camera position, subject state, motion blur/freeze where relevant · carry a 2–4 word burned-in caption (scene label, not subtitle).

• classicMD panels: smooth compositions, elegant typography zones, cinematic lighting.
• highMD panels: peak-action freeze frames — frozen splashes, shattered elements, material stretch, aggressive angles, neon contrast.

Prompt skeleton:

Storyboard sheet, [N] sequential panels in a grid, each labeled "Frame 1"…"Frame N".
Panel 1: [scene]. Panel 2: [scene]. … Panel N: [logo lock / brand name].
Each panel: [camera angle], [motion state], [mood/lighting]. Style: [cinematic / kinetic].
Consistent color palette throughout. Clean storyboard design, thin borders between panels.

Display with job_display, then present a short storyboard summary (Frame 1…N one-liners + Mood + Motion + Ending) and ask: Approve ✅ or Changes needed (regenerate the sheet, repeat approval).

STEP 4 — Generate the video

Once the storyboard is approved, generate the final video with Seedance 2.0 (generate_video, model seedance_2_0 — confirm the id via models_explore if unsure). Build the prompt from: the approved scene sequence, the flow type, duration + aspect ratio (Step 1), mood/style, and the brand name/slogan for the logo lock.

• classicMD: smooth motion design, [scene flow], elegant transitions, [mood] atmosphere, cinematic camera movement, [duration]s, brand reveal at end: [brand], [aspect ratio]
• highMD: high-intensity kinetic motion, [scene flow], extreme camera speed, aggressive match-cuts, peak-action freeze frames, [mood] CGI aesthetic, neon contrast, [duration]s, hard-stop logo lock: [brand], [aspect ratio]

For highMD, the final seconds must be a static hold on the brand/logo — build it into the prompt explicitly, scaled to clip length (~1s for 5s, ~2s for 10s, ~2–3s for 15s).

Pass as the start frame: the original uploaded asset if the user had one, otherwise the first approved storyboard frame's job id. Seedance 2.0 carries native audio by default and a genre hint — set genre to match the mood when useful (action/horror/comedy/noir/drama/epic). Display with job_display.

> Resolution note: Seedance 2.0 reaches 4K only in mode=std; mode=fast caps at 720p. See higgsfield-seedance / the specs layer.

STEP 5 — Review & iterate

Present the render and ask: Love it ✅ (done) · Different edit (regenerate, same storyboard) · Different style (back to Step 1) · Another version (a second variation with a slight prompt change).

Notes & rules

• Ask all Step 1 questions at once; never split intake.
• Storyboard = one grid image via GPT Image 2; never N separate images. No separate moodboard step — go brief → storyboard.
• Image model: GPT Image 2 (gpt_image_2). Video model: Seedance 2.0 (seedance_2_0).
• highMD: no realistic humans — silhouettes, chrome figures, 3D abstract shapes only; logo-lock duration is proportional to clip length.
• classicMD logo can appear as opener, closer, or both — ask if unspecified.
• Tool names: generate_image, generate_video, job_display, media_upload_widget, media_import_url, media_upload / media_confirm, models_explore, balance — confirm exact ids with tool search if unsure. Check credits with balance if the user seems concerned about usage.
• Match the user's language throughout. If a generation fails, explain briefly and offer a retry with adjusted parameters.

Related skills

• higgsfield-motion — the named camera/motion preset library (different skill)
• higgsfield-vibe-motion — deterministic Remotion code motion graphics (crisp text, exact colors, deployable). Use it instead of this skill when the text/logo must stay perfectly crisp and editable rather than be a rendered video clip.
• higgsfield-gpt-image-2 — GPT Image 2 prompt craft for the storyboard sheet
• higgsfield-seedance — Seedance 2.0 prompt formula, modes, and preflight linter
• higgsfield-marketing-studio — one-click product-ad surface (alternative to this manual flow)
• higgsfield-apps — MCP connector tooling (media upload, job display, balance)