ad-video

Ad Video

With motion graphics + AI-generated visuals, ad videois made. The render engine is HyperFrames (HTML→video), the visual source is the image-gen policy (Higgsfield CLI), the performance principle is A/B variants — even for video, pulling just one is spending without learning.

Prerequisites: hyperframes skill installation required (if absent, npx skills add heygen-com/hyperframes --all). This skill is a workflow orchestrator — render rules are owned by hyperframes.

Workflow

1. Brief (at once). What is being sold (product/URL), target placements (Reels/TikTok 9:16, feed 1:1, YouTube 16:9), length (default 30s, range 15–60s), hook candidates, brand tokens (DESIGN.md — if absent, brand-guide first), CTA. If you're going as far as ad delivery, budget/period is at the paid-ads gate — this skill is production only.

2. Routing (follow hyperframes rules).

• If there's a product/site → /product-launch-video (site capture + brand tokens)
• ≤10s cut where motion is the message → /motion-graphics
• Free composition with only assets → /general-video After routing npx hyperframes skills update <workflow> — as the router instructs.

3. AI visuals (image-gen policy as-is).

• If you need images for scenes, use the image-gen skill: via Higgsfield CLI, default gpt_image_2, no fallback, READ-verify the output. Do not bake text into images — all copy goes in HyperFrames typography layers.
• Product cutout: rembg + birefnet-portrait(people) — local, free.
• BGM/SFX/icons are handled with the media-use skill (royalty-free only).

4. A/B variants (MANDATORY). Per campaign at least 2, recommended 3 — with different axes:

• 훅 축: make the first 3 seconds different (number hook / problem statement / result first)
• Visual axis: product-centric vs situation-centric vs typography-centric Reuse shared scenes as subcompositions to lower variant cost. Naming {campaign}_{hook축}_{visual축}.mp4.

5. Ad spec check (blocking before distribution).

• Hook + brand exposure in the first 3 seconds (assuming sound-off viewing — key copy as on-screen text)
• Safe zone: for 9:16, no load-bearing elements in the top ~140px/bottom ~350px
• Korean copy line breaking: keep-all, do not break at particles (humanizer rule)
• CTA fixed in the last 3 seconds + logo end card
• File: H.264, platform specs (Meta ≤4GB/240min but in practice <100MB, Reels 9:16 ≤90s)
• After rendering, extract frames and READ: visually verify 4 frames — first frame/hook copy/CTA/end card

6. Handoff. Variant set + one-line axis description for each variant → user approval → delivery is paid-ads(budget gate), organic posting is organic-social. After 24–48h, based on performance, turn off the losing variant and make the next set with the winning axis.