Render Mosaic Grid Reveal Agent Skill

render-mosaic-grid-reveal

Render the mosaic-grid-reveal format from a config. The signature of this format is a

full-bleed N×N mosaic (default 3×3) of the brand's real product stills that **pops in

one tile at a time** (scatter order, ease-out-back "pop"), building the whole grid; the grid

then clears (scale-up + fade) and a clean end card builds line-by-line — the brand's

real wordmark, then a small sub-label, then the tagline, then a CTA. Silent by design (a

light music bed is muxed separately via create-music-elevenlabs). The hook is RANGE — nine

maximally-distinct variants show how much choice the brand offers.

Everything is a real-DOM HTML scene frame-stepped to PNG via Playwright and encoded with

FFmpeg, so the wordmark, tile captions, and CTA stay pixel-crisp (a video model would

smear them). No generative video, no i2v, no AI-rendered text. This capability is **FREE and

deterministic** — the only paid step of the format (the music bed) is a separate media cap.

Inputs (config.json)

• width, height, fps — canvas + frame rate (default 1080×1920, 30).
• wordmark — brand logo SVG (full <svg> or bare <path> markup) or PNG path. Real DOM,

recolored via palette.ink. Never AI-render it. wordmark_viewbox, wordmark_width tune it.

• palette — bg (warm off-white), accent + accent_deep (brand color), ink, cap.
• grid — cols/rows (default 3×3), inset (outer margin), gap.
• tiles — one per cell (length must equal cols*rows): image (real variant still),

name (caption), bg (soft pastel echoing the variant), ink (deep caption color).

• pop_order — scatter order across cells (default corners → center → edges).
• timing — grid_in0, cadence, pop, hold, clear (seconds).
• end_card — sub, tagline_top, tagline_bottom (bold), cta, url. Brand's own approved copy only.
• eyebrow — optional top kicker held during the grid.

scripts/config.example.json is a filled example (the Pair Eyewear "RANGE" worked example);

the tile image/wordmark paths are brand inputs bound by the orchestrator at remix time.

Run

python3 scripts/build_html.py --config config.json --out hyperframe.html   # emits scene, writes duration_sec back
python3 scripts/render.py --config config.json --html hyperframe.html --out master-silent.mp4

build_html.py computes the full timeline and writes duration_sec back into the config so

render.py (which reads dims/fps/duration from the config) frame-steps window.renderAt(t)

at fps and FFmpeg-encodes the silent master. Then mux the create-music-elevenlabs bed:

ffmpeg -i master-silent.mp4 -i bed.wav -map 0:v:0 -map 1:a:0 -c:v copy -c:a aac -shortest out.mp4.

Rules

• FILL the frame — tiles are full-bleed and edge-to-edge, NOT small cards floating in an

empty margin. Tiles pop in one at a time (scatter order), never all-at-once or static.

• Real wordmark + real product stills only. Never AI-render the logo or text.
• Copy slots are the brand's own approved lines — never invent claims, customers, results,

prices, or shipping speed.

• Pick the 9 most DISTINCT variants (color-wheel spread) — that IS the RANGE hook.