scenario-video-ads

Scenario Video Ads

Overview

This skill is the director layer: it turns one product shot and a brief into a finished ad, delegating mechanics to the sibling skills it names per stage. Connection and the core loop: see the scenario skill in this repo. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap. The full sibling map and install commands: references/dependencies.md.

Two rules bind every stage, because they carry the ad's credibility. The product is never generated: every product shot starts from the uploaded photo, as a first frame or composited into an approved still. Text is never generated: logos, taglines, prices, and legal lines are overlaid in assembly on text-free plates.

Ask the brief once, then run without stopping. The one approval that matters sits between the priced storyboard and generation: video is the expensive stage, so the user signs off on the budget table first. The concept page and the stills are cheaper checkpoints: show them when someone can answer. In a run where nobody can, continue: the brief's spend ceiling stands in for the sign-off, so generate video only while the priced board fits it, and the fidelity gate stands in for still approval.

Quick reference

Worked example: a perfume ad, 9:16, 20 seconds

1. Brief once: awareness, twist creativity, 20s for Reels and TikTok, ceiling in CU. Upload the product shot, then build the fidelity checklist from it with asset_analyze, instructing it to inventory label text, cap geometry, and glass tint: every later check reuses it (asset_describe returns a style line, not this inventory).
2. Concept names its anchor: open inside the gala-and-gown cliche, break it on beat three. Show the page to the user.
3. Storyboard 7 panels at 2 to 4 seconds each, hero shot designed first, 3 hook variants, 2 inserts as cut cover. dry_run every planned run, multiply by retry floors, present the budget table.
4. Stills at 9:16: composite the real bottle into styled plates with an image-edit model, baseline-plus-delta prompts. Each still is approved before animating, by the user or by the fidelity checklist when nobody can answer: image money is cheap, video money is not.
5. Animate each approved still: model_schema_get first, prompt movement first ("slow push-in over 4 seconds, the bottle rests fixed in place"), generateAudio: false when the schema lists it.
6. Gate every clip: frame-extract, then one asset_analyze call against the checklist. Label drift fails the clip; regenerate from the same still, never from the drifted output.
7. Music at a stated BPM so cuts land on bars; compose the timeline, scenario-text-overlay renders the tagline card into the safe zone; caption the master.
8. Pre-flight per delivery.md, then report spend from usage against the approved board.

Common mistakes

• Reaching for local ffmpeg or a desktop editor for the cut: assembly is model_run on Scenario tool models (see scenario-video-assembly).
• Letting the video model render a label, logo, tagline, or price: generated type drifts frame to frame; composite text in post.
• One long clip at the schema's maximum duration: generate 3 to 5 second product shots and 5 to 10 second coverage, then cut to rhythm in assembly.
• Judging fidelity by watching a clip once: drift hides between glances; run the frame-extract gate.
• Leaving the schema's audio toggle (generateAudio, where offered) at its default: the soundtrack is built in assembly, so clips with baked-in audio fight the bed.
• Cropping a 16:9 master to 9:16: each ratio is its own composition, decided at the storyboard.
• Two objectives in one brief: brand and performance in the same asset serves neither; make two cuts from one board instead.
• Shipping without the pre-flight: a sensual perfume cut or a wrong-length CTV file dies in ad review, not in generation.