Assemble an animated character-explainer video ad (~77s, 9:16, any art style: Pixar-style 3D, claymation, anime, paper-cut, felt) on an "N types of X" listicle spine: one recurring human protagonist plus one persona character per list item. Given per-scene clips, a scene timing table and narration, it trims, normalises, concatenates and mixes them, reusing the source ad's audio (restyle) or fresh per-scene voiceover over a music bed (original), loops a scene's keyframe when its clip failed, and burns captions last. Free, deterministic assembly (Python and ffmpeg, no paid calls); the recipe supplies the clips, keyframes, audio and captions. Use for the 3d-character-explainer format.
Install gooseworks: https://gooseworks.ai/creative-agent/install.md # then, in Claude Code, Cursor, or Codex: /gooseworks use the render-3d-character-explainer skill
The free, deterministic renderer for the 3d-character-explainer video ad format — an animated character spot (art style is the user's choice; the demo was glossy Pixar-style 3D) built on an "N types of X" listicle spine, where a recurring human protagonist plus a locked cast of N persona characters (one per list item) carry a hook → "deeper story" → cast-reveal → one beat per list item → kicker → product test → relieved payoff. This capability is the FREE assembly stage only. All generative work (Nano-Banana cast anchors + per-scene keyframes, Kling-V3 i2v clips, ElevenLabs VO + music, or a source ad's audio reused verbatim) happens upstream in the recipe and is handed to this capability as files.
It ports the validated compose recipe from the worked example — a "six types" Pixar-style 3D
listicle restyle (_render_full.sh — per-scene trim → normalize 1080×1920/fps30 → concat -c copy → mux the
source audio, with a static-still fallback on any failed clip). The assembly is
deterministic — iterate the cut for free, re-roll only the offending paid beat.
The recipe asks these of the user before any paid step; this capability assembles whatever comes back. None of them is fixed by the format.
visual_style — the art style of the anchors, keyframes and Kling motion prompts (Pixar-style
3D, claymation, anime, paper-cut, felt). Asked of the user; the demo used glossy Pixar-style 3D.narrator — who tells the story and in whose POV the lines / captions are written. Asked of
the user; the demo used an off-screen storyteller.narrator_voice — the VO voice (ORIGINAL mode only). Asked of the user; the demo used one
ElevenLabs narrator.tone — cute/comedic, sincere, educational, dramatic, punchy; shapes the style prefix, the
delivery and the script. Asked of the user; the demo was cute and comedic.protagonist — the human's age, gender, look and outfit (locked via the world lock). Asked of
the user; the demo used a young protagonist in a single tee.music — the bed under the VO, or none (ORIGINAL mode → music_bed). Asked of the user; the
demo used a light instrumental underscore.audio_mode: "restyle", the reference run) — re-tell a finished source
ad, beat for beat, as an animated character story in the chosen style and tone. The source ad's audio mix (VO + music bed) is
reused VERBATIM (source_audio), and the per-scene target_sec table is inherited from
the source's scene timing. No new VO or music is rendered. The trims must sum to the source
audio length.audio_mode: "original") — the ad authors its own narration. Each scene
carries a measured VO cue (scenes[].vo, target_sec = the ffprobe'd VO duration) which is
concatenated into a VO track (loudnorm I=-14) and optionally mixed under a music_bed
(loudnorm I=-26 then volume, amix normalize=0).target_sec and normalized to
identical dims/fps/SAR
(scale=W:H:force_original_aspect_ratio=decrease,pad=W:H:(ow-iw)/2:(oh-ih)/2:color=<pad>,fps=30,setsar=1).
A clip shorter than its window is extended with tpad=stop_mode=clone; a longer one is
-t trimmed. Decrease+pad (never crop) preserves the full 9:16 keyframe framing.clip is missing or failed to render, the
scene's keyframe PNG is looped (-loop 1) for target_sec, so the master always
assembles. Fallback scenes are printed at the end.libx264 -crf 18 -pix_fmt yuv420p -r 30 even if already correct — a dims/framerate mismatch makes the concat demuxer
silently drop frames — then concatenated via the concat demuxer (-c copy).source_audio muxed verbatim (-map 0:v -map 1:a), clamped to the
video length. Original: per-scene VO track (optional atempo, apad, -t clamp) → loudnorm
→ optionally mixed under the music bed.make_captions.py emits a libass .ass (one cue per scene, start = scene_start + 0.08s, suppressed on any scene with no caption — e.g. a product/end-card beat
carrying its own typeset copy). compose.py burns it as the final filter so captions sit on
top. Word-level energy-pop captions (Whisper on the narration) are the recipe's upstream
option — produce that .ass externally and point captions_ass at it; compose burns
whatever .ass it's handed.scripts/make_captions.py — emits the per-scene libass .ass from the SAME scene table
compose reads, so caption windows stay in lockstep with the cut. Run before compose.py
(or leave captions_ass unset / pointing at nothing to skip captions).scripts/compose.py — the assembler: per-scene trim + identical 1080×1920/30fps re-encode
(static-still fallback on missing clips) → concat → audio (restyle verbatim / original
mix) → burn captions → master mp4.scripts/config.example.json — the shape of the config the recipe binds (the
"six types of laundry stains" restyle demo, a made-up brand, as a worked example; captions, audio and
music are the demo's, never defaults).--config + a runtime work dir — NO hardcoded paths)config.json carries: audio_mode (restyle | original), scenes[] (each {id, clip, keyframe, target_sec, caption?, vo?, atempo?} where target_sec is the source-inherited
window in restyle mode or the measured VO window in original mode, and keyframe is the
static-still fallback source), source_audio (restyle), music_bed + music_volume +
atempo (original), width/height (default 1080×1920), pad_color (letterbox colour),
captions_ass, and caption_style. See config.example.json. pad_color defaults to neutral
black; set it from the brand palette.
generate_audio was false upstream — Kling would otherwise invent its own dialog track;
the real narration is muxed here separately. (This is the recipe's upstream call, not this
capability.)start = scene_start + 0.08s (avoids the caption flashing a frame before a cut).watch (QC the final master — the human protagonist reads as the SAME person every scene
(wardrobe/hair/lighting held), each persona is on-model, the cast-reveal lineup matches the N
list items, the product beat shows the REAL box, narration lands beat-for-beat, and duration is
within ±0.1s of the summed windows). The recipe gates the paid create-image-fal (cast
anchors + keyframes), create-video-fal (Kling-V3 i2v), create-vo-elevenlabs, and
create-music-elevenlabs calls to their own capabilities — this capability itself makes NO
paid calls.
visual_style — the art style of the anchors, keyframes and Kling motion prompts (Pixar-stylenarrator — who tells the story and in whose POV the lines / captions are written. Asked ofnarrator_voice — the VO voice (ORIGINAL mode only). Asked of the user; the demo used onetone — cute/comedic, sincere, educational, dramatic, punchy; shapes the style prefix, theprotagonist — the human's age, gender, look and outfit (locked via the world lock). Asked ofResearch 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.
A seven-step workflow for turning a concept brief into a finished animated explainer ad with Claude Code, image models and ffmpeg.