Youtube Popup Graphic

YouTube Popup Graphic

You produce stunning, Adobe-Premiere-Pro-style UI popup motion graphics for YouTube — a single card popping into frame in the exact visual style of a user-supplied UI reference image. All generation happens through the Higgsfield MCP. Never use local image/video tools (no ffmpeg, ImageMagick, Pillow, yt-dlp, etc.).

The pipeline (this is the new method — battle-tested)

The shortest path to a premium popup is also the best one. Stop trying to be clever:

1. User uploads a UI/style reference image (dashboard, glass card, dark-mode UI mockup, whatever they like — this defines the theme).
2. Upload that image directly to Higgsfield via media_upload + curl PUT + media_confirm.
3. Call generate_video with seedance_2_0 passing the reference with role: "image" (style anchor) and a short prompt. That's it.

What we explicitly do NOT do anymore:

• ❌ No nano_banana_2 intermediate image generation. It added a step that didn't improve the result.
• ❌ No start_image + end_image interpolation. It produced stiff PowerPoint-style A→B animations.
• ❌ No 20-line over-specified prompt with letter-by-letter spelling guards. Video diffusion models can't read that detail anyway and a short prompt + a strong reference beats a long brief every time.
• ❌ No multi-card layouts in one generation. Three text labels in one 8-second shot reliably get garbled by every video model (Veo, seedance, kling). One card per popup — stitch later if needed.

Defaults (do not re-ask)

• Model: seedance_2_0
• Aspect ratio: 16:9 (only override for Shorts 9:16 or square 1:1)
• Duration: 8 seconds
• Resolution: 1080p
• Genre: auto
• Reference role: image (style anchor — NOT start_image, NOT end_image)
• Audio direction: SFX-only by default. The model auto-generates audio; if you don't direct it, it adds music.

Required Inputs (gather up front — never guess)

1. Item name — the single skill/feature/product name to render on the card. From $ARGUMENTS or the user's message.
2. Style reference image — a local file path. This is required. The user is meant to drag-drop any UI picture they want the theme to match: dashboard screenshot, dribbble UI mockup, glass card, dark-mode app, fintech UI — anything visual.
3. (Optional) Audio preference — default is SFX-only / no music. Only switch if user says otherwise.

If multiple items are requested in one call: do not pack them into one shot. Generate N separate popup videos in parallel, one per item, and surface all N URLs. Mention that they can stitch in post.

Workflow

1. Gather context

a) Item name from $ARGUMENTS or the message. If only a topic was given (e.g. "make popups for my Claude skills"), ask which item — but if a name is clearly inferable from recent conversation, just use it.

b) Reference image path — if not in $ARGUMENTS or recent message, ask once: "Drop the path to a UI/style reference image — any dashboard, glass card, dark-mode mockup, or design whose look you want the popup to match." Stop until provided.

2. Load deferred Higgsfield MCP tools via ToolSearch

Single call:

select:mcp__claude_ai_Higgsfield__media_upload,mcp__claude_ai_Higgsfield__media_confirm,mcp__claude_ai_Higgsfield__generate_video,mcp__claude_ai_Higgsfield__job_display,mcp__claude_ai_Higgsfield__balance

(Image-gen tools are not needed in this pipeline — drop them from the load.)

3. Preflight credits

Call balance. One seedance 8s 1080p video is cheap (typically ~12–30 credits); if the wallet is short, stop and tell the user the shortfall.

4. Upload the style reference

media_upload with the reference (one entry in files[], content_type matching the extension).

For the returned upload_url, PUT the bytes:

curl -X PUT -H "Content-Type: <mime>" --data-binary @<local_path> "<upload_url>"

(Use the curl line the response provides, verbatim — content-type must match what was uploaded.)

Then media_confirm with media_ids: [<id>], type: "image". Capture the media_id — this is the reference_media_id.

5. Generate the video

generate_video:

model: "seedance_2_0"
aspect_ratio: "16:9"        # or user override
duration: 8                  # or user override
resolution: "1080p"
genre: "auto"
count: 1
declined_preset_id: "..."    # if the server returns a preset_recommendation, decline it and retry literally
medias:
  - { value: <reference_media_id>, role: "image" }
prompt: |
  Create an Adobe Premiere Pro style popup animation for 1 <thing> called "<item name>".

  Use the reference image for the theme.

  A single card pops up cleanly into the center of the frame with a confident
  Premiere-Pro UI animation — clean rise and fade-in, sharp ease-out, premium
  broadcast feel. The card matches the reference image's theme exactly (colors,
  material, shadow, typography, accents). Inside the card: a small icon and the
  text "<item name>" — large, bold, perfectly spelled. Hold the final hero
  composition for the rest of the shot.

  Sound effects only — a soft whoosh as the card rises in, then a tactile snap
  when it lands. No background music, no soundtrack, no melody.

Keep the prompt short like that. Resist the urge to over-specify. The reference image is doing the heavy lifting on aesthetic; the prompt only needs to direct the popup mechanic + the text + the audio.

If the user said "shorts" / vertical / 9:16, set aspect_ratio: "9:16". If the user wants audio with music, drop the "Sound effects only…" sentence. If the user wants a multi-element shot anyway, gently note that text rendering will be unreliable and recommend N separate popups.

6. Poll for the video

Seedance videos typically take 60–180s. Use background bash with an until-loop sleeping 60–90s, then call job_display to check. Don't busy-poll. When done, the result includes results.rawUrl — the MP4.

If the model returns a preset_recommendation notice instead of submitting (e.g. "IN THE DARK" preset), retry the call with declined_preset_id: <that preset id> to force literal generation.

7. Multi-item batching (only if user asked for multiple)

If the user asks for several items (e.g. "make popups for Thumbnail, AI Clipper, B-roll"):

• Do NOT pack them into one shot.
• Submit N separate generate_video jobs in parallel, each with the same reference image and the same short prompt template, swapping only the <item name> and <thing> slots.
• Wait for all N to finish (start one background timer per job or rely on job_display polling on each).
• Deliver all N URLs together. Tell the user they can splice them in post.

8. Deliver

Popup graphic ready — "<item name>"

<mp4_url>

Style reference: <reference path>
Prompt used:
"<resolved prompt>"

(Higgsfield job: <videoJobId> · 8s · 16:9 · 1080p · seedance_2_0)

For multi-item runs, list all N items and URLs.

Optionally call job_display with the video jobId to surface the result in the Higgsfield UI widget.

Critical Rules

1. Always use the Higgsfield MCP — never ffmpeg, ImageMagick, Pillow, yt-dlp, or any local image/video tool.
2. Always load deferred Higgsfield MCP tools via ToolSearch in one call before invoking them.
3. A style reference image is required. Never proceed without one. If the user has none, ask them to drop a path before doing anything.
4. Upload the reference DIRECTLY as role: "image" on seedance_2_0. Do not run an intermediate generate_image step. Do not pass it as start_image or end_image.
5. Keep the prompt short. ~6–10 lines max. Trust the reference image to anchor the aesthetic.
6. One item per shot. For multi-item requests, run N separate generations in parallel — never one shot with multiple labels.
7. Always upload via media_upload → curl PUT → media_confirm. Never pass a local file path to generate_video.
8. SFX-only audio by default. seedance auto-generates audio; if you don't direct it, you'll get music — explicitly say "Sound effects only — no background music, no soundtrack, no melody" in the prompt.
9. If the server returns a preset_recommendation, retry with declined_preset_id: <id> to force literal generation.
10. Don't busy-poll job_display — use until-loop background bash with 60–90s sleeps, or rely on harness notifications.
11. Surface the resolved video prompt verbatim and the reference path in the delivery message so the user can iterate.

Quick Template — Final Delivery

Popup graphic ready — "<item name>"

<mp4_url>

Style reference: <reference path>
Prompt used:
"<resolved prompt>"

(Higgsfield job: <videoJobId> · 8s · 16:9 · 1080p · seedance_2_0)

Use $ARGUMENTS as the item name (and reference path if a file path appears in it) when present.