render-flat-vector-explainer

Assembles a flat-vector product-routine explainer: one illustrated host character walks through a countable N-step routine (e.g. the demo's soil mix -> plant food -> leaf mist -> root tonic), one step per beat, each beat carrying a large corner numeral, a labelled chip + one-line tagline, and the step's real product photo, closing on an "N products" grid + brand CTA. It reads as a premium DTC explainer (Spotify/Anchor flat-vector lineage), not UGC.

This capability is documentation-grade. The format's source recipe is a documented recipe, not a runnable end-to-end app, so this capability ships the config schema (scripts/config.example.json), the field-to-script map (scripts/PIPELINE.md), and a README (scripts/README.md) describing the FREE assembly steps the agent runs by hand with ffmpeg + Remotion + PIL. The paid generative steps are separate capabilities the recipe orchestrates and gates.

Choices

Creative calls the user makes (the recipe's choices, asked in one round before any paid step). The demo's value is only an example, never the default. The flat-vector 2D style itself is the format, not a choice.

• host — who the illustrated host is (gender, age, look, outfit) -> character.anchor_prompt. Asked; the demo used a mid-30s woman with a short black bob in green overalls.
• setting — where the routine happens -> the character keyframe prompts. Asked; the demo used a sunny windowsill full of houseplants.
• tone — the script's tone -> VO lines + expressions. Asked; the demo used faintly comedic overwhelm -> satisfied payoff.
• narrator_voice — the VO voice -> voice.voice_id. Asked; the demo used ElevenLabs "Eryn".
• music — the bed -> music.prompt (or none). Asked; the demo used lo-fi pop, 95-105 BPM.

scripts/config.example.json is the Greenmoth worked example: copy its structure, never its creative values or its brand palette.

The two non-negotiable separations

1. Motion layer != text layer. Animate a text-stripped clean plate with Kling i2v (subtle motion, style-preserving negative, cfg 0.5), then composite every chip / numeral / tagline / slate / CTA as an animated Remotion DOM overlay on top. Baking text into the keyframe before i2v warps the type and forfeits the ability to retime/restyle it — this separation is the format's whole credibility.
2. Real assets != AI assets. The per-step product photo and the closing "N products" grid are real product webps composited with PIL (AI duplicates SKUs in a grid). Only the character vignettes and stylized backgrounds are generative.

Free assembly steps (this capability)

The agent runs these deterministic, free steps by hand — see scripts/README.md for the ffmpeg/Remotion/PIL detail:

• Remotion overlay — import each Kling clip as the moving base; composite chips / numerals / taglines / slate / grid / CTA as animated DOM on top -> the animated silent master. Slate/grid/CTA beats are Remotion text with no i2v.
• PIL product grid — composite the N real product webps on the brand ground for the closing lockup; preserve each aspect (never stretch, never AI-dupe).
• Captions — word-by-word burned from the eleven_v3 with-timestamps char timings (libass); suppress on slate/grid/CTA scenes so two text layers don't collide.
• Audio mix + master — place each VO line at its scene start, duck the music under VO (sidechaincompress), loudnorm I=-15 VO-forward, mux, burn captions LAST -> finals/master-final.mp4 (~50s).
• 30s cut — slice each beat's region OUT of the animated silent master (never a static intermediate); trim short beats, gently slow long beats (setpts <=1.6x), re-burn scaled captions -> finals/master-final-30s-v1.mp4.

Paid gen steps (separate capabilities)

The recipe orchestrates and gates these; they are not part of this capability:

• Flat-vector character anchor + per-scene keyframes + clean plates -> create-image-fal (nano-banana; re-render a FRESH flat-vector anchor, never chain a photoreal ref).
• Kling i2v on the character scenes -> create-video-fal (Kling v3 standard image-to-video with start_image_url and generate_audio: false, cfg 0.5; Kling 2.5 Turbo is withdrawn, style-preserving negative, low motion; TEST one scene before batching).
• Full-sentence VO -> create-vo-elevenlabs (eleven_v3, with-timestamps).
• Music bed in the chosen style (demo: lo-fi pop; skip if the user chose no music) -> create-music-elevenlabs.

Contract

• Documentation-grade + FREE assembly (Remotion + PIL + FFmpeg); no paid calls in this capability, no AI-rendered text.
• Text is an overlay, never baked. Strip to a clean plate -> i2v -> composite text as Remotion DOM.
• Any multi-SKU grid is PIL of the real product webps; preserve each aspect ratio.
• Kling holds the 2D flat-vector look only at LOW motion (cfg 0.5 + style-preserving negative). Aggressive motion drifts to photoreal.
• Cut down from the ANIMATED master, never a static intermediate; frame-diff to prove localized motion.
• The paid steps — keyframes/clean plates, Kling i2v, VO, music — are separate capabilities (create-image-fal, create-video-fal, create-vo-elevenlabs, create-music-elevenlabs); the recipe orchestrates them and gates the spend.