Assemble a split-screen creator ad from a config: an AI-creator lip-sync take fills the bottom half while real 16:9 product or demo clips play uncropped in the top half over a blurred fill of the same clip (never black bars), with a brand-colour divider between them. Scenes follow the voiceover timing and are joined hard, an end card holds about 3 seconds, then the finished cut is transcribed and word-level captions are burned in the chosen style. Free, deterministic assembly and captions; the voiceover comes from create-vo-elevenlabs, the creator anchor from create-image-gpt-image-fal, and the lip-sync take is supplied as an input. Use for the split-screen-creator format.
Install gooseworks: https://gooseworks.ai/creative-agent/install.md # then, in Claude Code, Cursor, or Codex: /gooseworks use the render-split-screen-creator skill
Assemble a split-screen creator ad from a config: a two-zone vertical (1080×1920, 9:16, ~40s) format where an AI creator talking-head anchors the BOTTOM ~48% of the frame and real 16:9 product/demo clips run uncropped in the TOP ~52%, a 3px brand-color divider between the zones. The creator delivers the whole VO cold-to-camera and each top clip proves the claim its VO line makes. This capability is the FREE, deterministic assembly + captions — the two-zone composite (contain-fit + blurred-cover fill + divider + creator slice), the hard-concat, the end card, and the word-level caption burn from the assembled cut.
scripts/config.example.json is one worked example (the fictional
Larderloop "whole Sunday" demo, ~40s 1080×1920 9:16, 6 scenes + an end card) — copy its
structure, never its creative values;
scripts/PIPELINE.md maps every config block to its source step and
scripts/README.md documents the free assembly.
The creative calls are made upstream by the user (the calling format's recipe
choices) and arrive here as config. This capability never picks them.
creator_gender / creator_age / creator_look — who the creator is
(creator.brief); shapes the anchor and the voice upstream. Here it only
affects framing checks. The worked example used a woman in her 30s.setting — where the creator films from (creator.brief.vibe_setting).
The worked example used a lived-in kitchen with window daylight.tone — the script and VO delivery; upstream only. The worked example
was a confident, fast explainer.caption_style — captions.style (serif-accent, kinetic-pop,
neon-glow, clean-bubble). Burned here. The worked example shipped
serif-accent.If a creative field is empty in the config, ask for it — don't fall back to the worked example's value.
This is the FREE, deterministic assembly + captions stage — it spends
nothing. The paid inputs are separate steps — the VO (create-vo-elevenlabs,
ElevenLabs eleven_v3 with-timestamps, sliced into per-scene windows), the
photoreal MEDIUM chest-up AI-creator anchor (create-image-gpt-image-fal,
gpt-image-2) — shot at a natural webcam distance (headroom + shoulders, real
room), not a plain-background close-up headshot (see the anchor note below), and
the whole-VO lip-sync (a paid VEED Fabric 1.0 @ 720p take — a no-atom step,
image_url = the anchor, audio_url = the vo mp3, billed per second of VO and the largest spend; run its calls sequentially, veed/fabric-1.0 storage-auths 403 under
parallel load). Given the creator lip-sync take + the per-scene VO timing + one
16:9 top clip per scene + the scene-1 hook graphic + the end-card clip,
render-split-screen-creator composites the two zones, hard-concats the scenes,
appends the end card, transcribes the assembled cut, and burns the captions → the
master. Re-cuts reuse the existing VO / lip-sync / clips and spend nothing.
top_height ~998. Keep every stacked height EVEN
(998 + 4 divider + 918 = 1920) — libx264 rejects odd dimensions.ugc-walk-and-talk). VEED Fabric handles photoreal fine (unlike Seedance).top_start/top_end) to the on-message segment that proves its VO line.
Never loop a short clip — set the window and the assembler speed-fits it to
the scene (looping replays into a sparse/black tail).timing.json; the
lip-sync drives the mouth.endcard.clip_end),
not the black tail.captions.style (serif-accent, kinetic-pop, …). Keep the -precaption cut + the
.ass sidecar so captions restyle without re-rendering the composite. If the
host ffmpeg lacks libass, render the cues as timed PIL PNG overlays (ffmpeg
overlay=…:enable='between(t,st,en)') at the same placement.loudnorm I=-14 → a
1080×1920 h264+aac master. No paid calls, no keys — the VEED Fabric lip-sync is
a supplied input, produced upstream.creator_gender / creator_age / creator_look* — who the creator issetting* — where the creator films from (creator.brief.vibe_setting).tone* — the script and VO delivery; upstream only. The worked examplecaption_style* — captions.style (serif-accent, kinetic-pop,Research supported advertising promises from a specific buyer's context, current product facts, customer language and observed ads. Turn them into static, video or copy ideas; match video ideas to actual recipes and save a product-scoped angle bank for write-video-ad-script. Product-led discovery can use verified facts without reviews. Reference persistence and reach are observations, not proof of performance.
Analyze ad campaign performance data (Google, Meta, LinkedIn) to identify what's working, what's wasting budget, and specific cut/scale/test recommendations. Runs statistical analysis, funnel diagnostics, and multi-channel budget reallocation with specific dollar-amount shift recommendations and scenario modeling.
For paid lead-gen and participant-recruitment ads, replaces vanity CPA with true CAC per qualified lead by joining ad-platform data with downstream funnel events, surfaces tracking gaps, and classifies every creative into Scale / Keep / Investigate / Cut.