Scenario Ugc

Scenario UGC Creator Video

Overview

UGC is a register, not a length: content that reads as a person talking into their own phone, not a brand talking through a camera crew. Everything in this skill serves that register, and most failures come from importing ad craft into it. This skill routes the production; mechanics live in the sibling skills named per lane. 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.

Two rules are non-negotiable. The product is never generated: demo shots start from an uploaded photo or footage of the real product (scenario-product-shots for stills). And the words are never invented: no fabricated testimonials, review counts, metrics, medical or financial claims, or legal copy; speak only lines the user supplied or approved, and prefer observable statements ("the texture looks lighter") over claims ("this cures acne").

Quick reference: route by speaker lane

Discover members with recommend, passing the lane's capability and the user's own words: it ranks by measured cost and latency and names the purpose-built pick, where a capability-worded search returns hundreds of keyword hits with nothing to choose between them. Read next_step before taking a pick, per the scenario skill. Keep search for a member you can already name. Never assert a generative model's id as a constant. Scenario's own single-purpose tool models are named outright below: there is exactly one of each, so discovering them would only re-derive a constant.

Script in six spoken beats: hook (one concrete tension or result), context (why this speaker cares), product moment, proof (visible demo or a user-supplied fact), turn (objection answered or before-after), close (soft CTA). Write for the mouth, not the page: contractions, false starts allowed, no taglines. Spoken pace runs near 2.5 words a second, so a 25-second ad is roughly 60 words, but that is a first guess and avatar members undershoot it: observed rates run 1.76 to 1.94 words a second, and one member padded a 60-word script with 21s of silence instead. Time the returned clip before assembling, then trim the silences where there are any and cut the script where the delivery is simply slow. The compositor has no cut inside a layer, so a trim is one layer per kept segment, each with its own trimStart and duration.

Keep the register in every visual prompt: phone-height framing, available light, a real location with clutter, natural skin texture, one handheld drift at most. Cinematic grammar (dolly moves, golden-hour rim light, shallow anamorphic looks, graded color, retouched skin) reads as an ad and kills belief. Compose 9:16 natively, reframing any non-vertical still a lane consumes (the avatar portrait, a motion-control character still, the product photo) per scenario-formats before generating; a cropped 16:9 master frames like television.

Worked example: 25-second founder ad from a portrait

1. Brief once: offer, platform, runtime, the facts the founder may claim, tone. Collect the portrait and the real product photo, then run without stopping.
2. Create this run's collection before the first generation (collection_create, catalog write lane, name only), then collection_add_assets each keeper as it lands; its returned itemCount is the receipt.
3. Draft the six beats at about 60 words and confirm the wording with the user; the script is a claims surface, not just copy. Unattended, keep every line to wording the brief already supplied and cut any beat that would need a new claim.
4. Voice: upload_asset the founder's recorded narration, or generate TTS per scenario-elevenlabs when they want a stand-in voice they approved.
5. recommend with capability="img2video" and the brief in the user's words; pick from its ranked list by input contract (a script-taking member when the speech exists only as text, an audio-taking one when a recording exists), then by whether the member can be told not to burn in captions: some hallucinate gibberish subtitles that no negative prompt suppresses, and the ones with a switch cost several times more, which recommend prices for you.
6. upload_asset the portrait. Prompt the speaker's hands empty and the set free of products: avatar members invent props and label type. model_run with dry_run=true first: avatar members sit far apart on price, so quote before spending.
7. Run for real with wait=false, then jobs_wait with the returned job id, re-called with pending_job_ids on timeout, never a second model_run.
8. Demo insert: animate the uploaded product photo with an image-to-video member (scenario-video), 3 to 5 seconds, one micro-move. Whether aspectRatio survives a start image is per member and the schema note can be wrong either way, so read width and height off the returned asset rather than assuming the ratio held.
9. Assemble with model_scenario-compose-video per scenario-video-assembly: talking head as the spine, insert cut over beats three and four, product-name card from scenario-text-overlay as an image layer, dropped outright when no approved product name exists, since inventing one is the fabrication the brief forbids. Every overlay layer needs an explicit width and height or the compositor scales it, and durationMode: "custom" pins the master's length against image layers that would stretch it. A succeeded compose job is not proof the layers drew or landed where sent: pull a frame from the composite and confirm the card is present and placed before delivering. Captions come last, on the finished cut, from the captioner scenario-video-assembly discovers with recommend (two exist, so no fixed id), positioned inside the platform's safe zone: the compositor has no text layer.
10. asset_display the master. Gate every generated clip, the talking head included: take the frames from the platform, the free firstFrame and lastFrame off asset_get first and model_scenario-video-to-image-seq when a mid-clip frame is what settles it, then verify them against the uploads (scenario-asset-analysis); a local extraction is not a substitute at any budget, because it yields no asset to file or audit and the gate stops being traceable; an invented product in the speaker's hands or legible generated type fails a clip exactly like label drift on the insert. Report spend by summing this run's own job records by job id: jobs_list is project-scoped and over-reports, and usage's headline figure is project-lifetime.