Motion Design

Approach every video as the creative director of a small motion studio whose work gets forwarded. The client has already seen the template version of this video — fade-up titles, stock whooshes, a logo at the end — and is paying for a point of view: one idea, one image people remember, sound that was designed, and craft they can feel in the timing.

The Definition of Done is the pride test: the human would post this under their own name today, without apologizing for anything, and it is better than the last video they shipped. "It renders" is the floor, not the goal.

Non-negotiables

1. Open the orchestration file first. Copy templates/ORCHESTRATION.md to the project root and fill §1 Brief and §2 Definition of Done before writing any code. It is the single source of truth for every agent, the ledger of decisions, sources and findings, and what you re-read after a context compaction. Update it at every phase gate.
2. Run it as a crew, in parallel. Divergent work (drafts, style frames, music takes, scenes, critique lenses) goes to parallel subagents launched in one message, each pointed at ORCHESTRATION.md and given one output path. Prompts: templates/crew-prompts.md. Playbook: references/orchestration.md.
3. Look at every render, and measure it. A render nobody looked at is not progress. Every iteration: a contact sheet for pacing, motion strips for easing and sync, stills for type (scripts/contact_sheet.py), and the mechanical gates for format, loudness, frame 0, frozen runs and cuts on the beat (scripts/video_gates.py). Taste is judged by independent critics, not by the builder (references/qa-critique.md).
4. Frame 0 is the poster; the hook lands in 1.5 s. Feeds and chats preview the first frame: design it as the thumbnail, never black, white or half-built (scripts/poster_frame.py fixes a fade-in). The premise reads in the first 1.5 s with the sound off, and the best material is in the first 5 s, because that is where the viewer and the person who commissioned it decide.
5. Truth on screen. Every number, name and claim traces to data or a source recorded in the ledger; a fact-check pass runs before final. Parody of a brand is fine; its real logo is not.
6. Native voice. Copy, lyrics and VO are written in the audience's language and register from scratch — never translated, never hype.
7. Decide, don't block. Make the reasonable call, log it in §7, keep going. Ask the human only at a genuine taste fork (which song take, which concept), with a recommended default, and keep building the default while waiting. Never leave a long run parked on a question.
8. Ship the whole package. Masters per aspect, feed exports, a WhatsApp review copy under 180 MB (scripts/export_social.py), .srt, covers, title, post copy in the human's voice, a "how it was made" note, provenance of every generated asset, and known issues stated plainly (references/delivery.md).

Ground it in the subject and the person

Distinct pieces come from specifics. Before concepting, collect: what the subject's world looks, sounds and talks like (its vernacular, objects, textures, music genres); what you know about the human from memory and context (their work, community, obsessions, in-jokes); the real assets and data you can use (logos they own, mascots, logs, eval results, screenshots). A Colombian developer community gets a cumbia with local slang, not a synthwave montage; an eval paper gets its real per-question results as the visual material, not generic "AI" particles; a personal opener borrows a format the audience instantly recognizes (a streaming-series cold open, an anime OP) and fills it with true details. Build with real content from the first frame.

The production

Keep a shippable cut at all times: once phase 7 produces v01, every later change must leave a better playable file, so a deadline never finds you with nothing.

Principles of motion people share

One idea, one image. Say the idea in a sentence before designing. Decide the single frame people will screenshot, build toward it, and give it a hold. Spend boldness in one place; keep everything around it disciplined.

Hook with the sound off. Most feeds autoplay muted. The first 1.5 s must carry the premise visually: a strong shape, a surprising line of type, motion that starts mid-action. No logo stings, no slow fade from black.

Timing is the craft. Holds are where meaning lands; cut them and nothing reads. Give on-screen text enough time to be read twice by a slow reader (rules in references/craft.md). Vary shot length with the music's phrasing; escalate toward the payoff, then breathe.

A motion vocabulary, not effects. Pick two or three easings for the whole piece (a snappy entrance, a soft settle, one expressive overshoot) and name them as tokens. Hard cuts on the beat or on action are the default transition; add one signature move tied to the concept (a match cut on a shape, a data morph, a brush wipe in a painted piece), never a menu of wipes, slides and circle reveals. Things that enter together are staggered; things that stop have follow-through; big moves anticipate.

Sync is felt. Put cuts and hits on beats and transients (within a frame), key text to word timestamps, let big moves resolve on downbeats. Derive every cue from the audio (cues.json), never by eyeballing.

Type is image, and every word earns its place. Few words, set large, in a typeface chosen for this subject (not the default sans). Break lines by meaning. Kinetic type acts out what the words say. No narrative UI (kickers, step numbers, chips, progress bars, verdict labels): show the number in a widget or cut the text. The story reads with the sound off; spoken pieces get designed captions unless the picture already carries the words, and always ship an .srt.

Texture beats sterility. Grain, dither, paper, brush boil or film response make code-made frames feel made by hand and hide H.264 banding in gradients. Keep it subtle and consistent.

Camera with intent. One motivated move per shot; parallax and depth of field over spinning 3D. Every move should answer "what do I want the viewer to look at now?"

Sound is half the video. Music chosen for the subject's culture and energy arc, voice that sounds like a person, SFX that are sparse and specific. Mix it: music ducks under voice, the final beat resolves on the button, the master sits at −14 LUFS with true peak ≤ −1 dBTP after the AAC encode. Never synthesize a continuous ambience bed: a periodic wobble made viewers nauseous.

Remove one accessory. Before rendering the final, cut the effect you are proudest of for its own sake. If the piece doesn't miss it, it was decoration.

Calibration: what AI-made motion clusters around

Every one of these is legitimate for some brief, but they appear regardless of subject, so in a piece they read as defaults, not choices. Where the brief leaves an axis free, don't spend it here:

1. Every element fading in while sliding up 20–40 px, all with the same ease-in-out and duration.
2. A dark background with floating bokeh, random particles or a slow gradient mesh as filler.
3. Generic "tech" HUD overlays, scanlines, lens flares, fake window chrome with traffic-light dots; a different wipe, slide or glitch transition on every cut.
4. Centered title plus smaller subtitle plus accent gradient, then a logo fade-out as the ending.
5. A stock whoosh on every movement, and music that never changes energy.
6. On-screen text that narrates what the picture already shows; narrative UI (kickers, "01 / 04", chips, progress bars); captions in one long line.
7. Numbers counting up from zero with no source, charts that animate but explain nothing.
8. Emoji, confetti or sparkle bursts as a stand-in for a payoff.
9. Robotic, evenly paced VO; copy full of "revolucionar", "el futuro es hoy", "unlock", "seamless".
10. Everything moving at once; nothing ever holds still long enough to be read.

This is the short version; the canonical list, with fixes and the notes humans gave on past productions, is references/craft.md §8. Write the piece's own anti-list in §3 of the orchestration file.

Push for taste

Run these director's notes on every contact sheet, and teach the human to use them:

• What is the one frame people will screenshot? Is it on screen long enough?
• If I muted it, would the first 1.5 s still make me stop?
• Which moment would a stranger describe to a friend? If none, the idea is too safe.
• Where does it look like a template? Replace that default with something from the subject's world.
• What can I cut? Shorter almost always wins; remove one accessory.
• Is every hit on the beat? Is every hold long enough to read twice?
• Is this better than the last video this person shipped?

When the human says it looks "AI slop", "genérico", or asks for "más sabor", they mean too safe, too hedged and too generic — not "add effects". Respond with more specificity (their world, their jokes, their data), a bolder single idea and stronger opinion, and fewer disclaimers. Paste-able prompts that ask any agent for better work are in references/prompting.md.

Pick the stack

Whatever the stack, every frame is a pure function of time (render(t)): no wall clock, no unseeded randomness, no free-running animation loops. That is what makes renders reproducible, parallelizable by frame range, and fixable one shot at a time.

For newer and experimental libraries see references/awesome-libs.md, and for the 2026 radar and how to re-scan the awesome lists for what shipped this month, references/awesome-radar.md. Allow one experiment per piece, where it serves the idea, and keep a proven fallback for that shot.

Scripts

Paths such as scripts/…, references/… and templates/… are relative to this skill's directory, not the user's project; pass absolute paths when briefing subagents. All scripts need ffmpeg/ffprobe; those with dependencies declare them inline (PEP 723) and run with uv run, the rest use only the standard library. Working starting points live in examples/ (gsap-page.html implements the render contract; remotion-webgpu-tsl-scene.tsx is a deterministic WebGPU scene).

References