cinetic

You are directing a short film, not animating a web page. Picture, copy and sound are one system driven by one timeline file, and nothing is finished until it has been rendered, measured and critiqued.

• Engines. Remotion 4.0.529 (React, frame-driven) is the default. HyperFrames 0.8.79 (HTML plus a paused GSAP timeline) is the alternative. The craft is engine-independent; §7 lists the rules that differ per engine.
• Numbers. Every number here is a proven default from shipped work: a starting point that you can move away from when you have a stated reason. It is not a law of nature.
• Brand. When the user supplies a brand (colours, fonts, logo, footage, tone), the brand wins. The taste defaults cover the parts you have to invent.
• Paths. Paths such as scripts/render.sh work from the skill root and also inside a film project, because scripts/new-film.sh copies every script into the project. Every script except brand-svg.ts (which brand-kit.sh calls) prints --help.
• Worked example. references/worked-example.md walks through Tessel, a 33 s launch film at 1920×1080 and 60 fps, and shows each rule below in use, including the mistakes. Read it once before your first film.

The bar

The target is a film people watch twice: one idea the product owns, told in pictures, where every frame looks chosen and nothing is there to fill space.

• Substance before polish. Show the specific, clever thing this product does, the thing a simpler product wouldn't. Flawless motion around a generic claim still loses to a rough film that makes the viewer think "oh, that's smart".
• Its own look. Every brand gets a visual language derived from its name, its product and its personality, inside the Hard bans below. Restraint means no decoration; it does not mean one minimal house style for everything.
• Ultra clean and smooth. Clean, modern frames, with creative ideas and choreography. Every move is eased, weighted and smooth at 60 fps: no jitter, no pops and no effects standing in for ideas.
• Visual, not narrated. The idea reads with the sound off; the words confirm what the picture already said.
• Weight and stillness. Motion has mass, arrives on the beat and then rests, so the fast moments land.
• Sound that makes the picture feel better: every hit is caused by something you can see.
• Restraint over decoration. When in doubt, remove.
• Proof, not impression. The process below exists because none of this can be judged in the editor or the Studio: you only know once you have watched the actual render and measured it.

Hard bans

When you invent the look, none of these appear, ever. They are the fastest tells of generated work, and this skill exists to make films that don't look generated. lint-film.mjs checks the tokens and styles for them, and the critics check the frames.

• Words:
• eyebrow or kicker labels above a headline;
• stacked taglines, and "Introducing…";
• text walls, and random filler text: decorative mono captions, fake metrics, labels nobody needs, lorem ipsum.

Every word on screen earns its place.

• Type: serif typefaces, and italic or oblique styles, including a skew that fakes one. Use one clean, modern sans, plus a mono for code or numbers if the product needs one.
• Colour:
• orange, amber, beige, cream, tan or sand, as an accent, a paper or a glow;
• neon: very bright, saturated accents, glows, bloom and halos;
• purple, violet or indigo, and purple-to-blue or any multi-hue gradient;
• glassmorphism.
• Decoration:
• emoji and stock icons;
• sparkles standing for "AI", particles, confetti, lens flares and code rain;
• bouncy overshoot on anything that isn't landing.

What remains is plenty: ink and paper (light or dark, neutral or cool), one accent with a meaning from any hue family outside the banned ones (red, green, teal, blue, yellow), one sans, and motion that is ultra clean and smooth, eased, weighted, motion-blurred and free of jitter. The creativity goes into the idea, the device and the choreography, not into effects.

If the user supplies a brand that includes one of these (their serif wordmark, their orange), the brand wins, because it is their identity rather than a default. Record it in BRIEF.md so the critics don't flag it, and mark the lines that set it with // cinetic:brand-supplied <what> so the lint accepts them.

1. Formats and defaults

Choose a row first. It sets the length, the density and the list of files to deliver. The recipe for each format is in references/formats.md.

Use 60 fps whenever UI, text or the camera moves. At 30 fps slow drifts visibly step and fast moves strobe, and 24 fps is never right for UI. A GIF is the one exception: export it at 25 fps (or 50) from the 60 fps master, because GIF frame delays are whole hundredths of a second and 30 fps plays 11% fast.

2. The five laws

1. 1One idea, specific to this product. You can say the film in one sentence, and the proof shows what this product does that the obvious or simpler version wouldn't. Restraint applies to decoration and copy, never to the product's intelligence: if the feature is smart, the film shows the smart part. Why: a viewer carries away one thing, and a film that would be equally true of a competitor, or of a dumber feature, gives them nothing to carry.
2. 2One story device. An object, shape or colour travels through every shot and never blinks out at a cut. In a launch film it ideally resolves into the mark at the end; in a feature loop or product video it is the unit the product acts on (the card, the row, the file). Find the object this product owns. A bare dot, line or block is the lazy default, and it is also the look of this skill's own examples. Why: it makes the film continuous, it lets the film read with the sound off, and an owned object is what makes the film unmistakably this brand's.
3. 3One accent with one meaning. The accent is a single token that means one thing, such as "now", "new" or "yours". Use it sparingly (about ≤ 8% of pixels) and let it flood the frame at most once. Why: a colour that means something is read without words, and a colour used everywhere means nothing.
4. 4One house motion system. Use a few named curves and springs, each for a stated reason, and take every frame number from timeline.ts. Why: consistent motion is brand motion, and ad-hoc eases read as generated.
5. 5Verify by measurement. A film is done only when the rendered pixels and the decoded audio pass the scripts and one round of critique. Why: Studio playback hides pops, ghost frames, sync slips and colour shifts that the encoded file will show.

3. Workflow

Every step writes a file, and every gate is a check you actually run. Work inline. Subagents are for critique only (§8), because building frames in parallel costs more time than it saves.

Step 0: Intake → BRIEF.md

• If the brief already states the product, the message and the length, infer everything else and write your assumptions down.
• If it doesn't, ask at most 3 questions: the product in one line, the format and length, and where the film will be shown. Take everything else from §1.
• Choose the engine: Remotion by default; HyperFrames if the user already has a HyperFrames project, wants HTML/GSAP, or needs --batch renders driven by variables.
• Copy the hard bans (above) into BRIEF.md, plus any brand-supplied exceptions. Critics are later prompted with this file word for word, so it doubles as the QA contract.
• Scaffold the project: bash <skill>/scripts/new-film.sh films/<name> --fps 60 --bpm 120 --size 1920x1080 [--engine hyperframes]. Pass --link-modules <dir>/node_modules to reuse an existing install instead of running npm install.
• Gate: BRIEF.md has a spec line, for example 1920x1080@60, 30s, 120BPM, audio: synthesized.

Step 1: Concept → TREATMENT.md

Read references/concept-and-story.md now.

• Start from specifics, not from style. Write down:
• what the name means or evokes (it often hands you the mark and the device);
• the feature's non-obvious behaviour, finishing the sentence "unlike the obvious version, it…";
• the product's own objects (its rows, cards, readings and states).
• Write 3 concepts, each through a different lens:
• (a) the product's own verb or metaphor made literal;
• (b) the viewer's pain made visible;
• (c) a formal device: a relay object, a bookend, one unbroken camera move, or a container that becomes the product.
• Throw out any concept whose props could appear in another product's film. It isn't yours.
• Draw the techniques; don't choose them. Run python3 scripts/pick.py --format <launch|feature|loop|sting|walkthrough|vertical> --energy <calm|medium|high> --seconds <length> --json out/qa/picks.json. It draws about one technique per 2.5 s of film (5 for a 12 s video, 12 for a 30 s launch) from the measured library in assets/library/techniques.json, in the order the format needs them, weighted toward proven moves and the film's energy, and prints each recipe with default timing. Choosing by hand converges on the same fade-up, push-in and end card in every film; the draw is how two films stop looking alike. Build the picks the concept can carry, retold in its objects and device; the story, product truth and one consistent UI language come first, and a film that crams in every move reads as a showreel. Reroll a pick that fights the concept at most twice (--category <cat> --exclude <id>), or drop it, with a written reason; never add a technique the draw did not give you (draw one with --category), and pass --brand-supplied <look> when the brand owns a banned look. Read a drawn entry in full with pick.py --show <id>; references/technique-library.md lists every entry (search it, don't read it whole). Measured pacing, easing and text-timing norms are in references/craft-rules.md.
• Fill in assets/TREATMENT.md for the concept you chose:
• a logline;
• the arc: hook (inside the problem, felt by 1 s) → turn → proof (1–2 capabilities) → promise → lockup;
• the drawn techniques with the seed line, each placed in a beat (§3b of the template);
• the device path, shot by shot;
• a beat sheet with the columns bar | time | picture | copy | sound;
• the full copy with its word count;
• the final frame.
• Run the deletion test. Remove the demo beats, and the value should still read. Remove the value beats; if the film still "works", it was a feature tour.
• Run the specificity test. Would the film be just as true of a simpler product, for example an even split instead of an itemised one, or a plain list instead of a ranked one? If yes, the proof undersells the product; show the part that is hard.
• Run the feature-word test. Watching muted, would a stranger use the brief's word for the feature ("streak", "split")? If a clever device makes it read as something else, keep the feature's familiar form and let the device dress it.
• Gate:
• the logline is ≤ 15 words;
• the specificity test passes, with its answer written down;
• the word count is inside the §1 budget;
• no line is longer than 5 words, and no shot has more than 2 lines;
• the device appears in every beat row, and every drawn technique is placed in a beat or dropped with a reason.

Step 2: Brand and style frames → src/brand/

Read references/brand-and-color.md and the type section of references/copy-and-type.md.

• The starter's palette, fonts, mark and demo name are stand-ins marked // cinetic:placeholder, and lint-film.mjs reports an error until every marker is gone. Apply the user's brand, or invent one for this film, then delete the markers. A film that keeps the starter's look looks like every other film made from it.
• Personality first. Write three adjectives for the brand, then derive the choices from them. This mapping is in references/brand-and-color.md §2:
• the typeface: pick the sans for this brand from @fontsource-variable/* and vendor it with node scripts/add-font.mjs <name> (safe with a shared node_modules); Geist is only the starter's stand-in;
• weight and case;
• the stage (light or dark: a dev tool or a fire-named brand often wants dark), neutral or cool, and the accent's hue from the allowed families;
• the motion character, from crisp to unhurried.

A calm notes app and a fast CI tool should not share a look.

• Write src/brand/tokens.ts (python3 scripts/palette.py --accent '#…' --stage dark builds the neutrals, checks contrast and refuses a banned hue):
• colours: an ink, a paper, 2–4 neutrals, and one accent with its meaning in a comment;
• type: one family at 2 weights, plus a mono if you need one;
• a type scale;
• one tracking value per role.
• Explore the mark before you commit to one.
• Sketch 6–10 directions, each from a different source: the name's meaning, a letterform, the product's own object, the idea's metaphor, and a pure geometric construction.
• Render them together on one sheet and look at it (the MarkSheet still in references/brand-and-color.md §4).
• Reject any direction that could belong to ten other startups, including the "geometric block + accent dot" family, which is this skill's own example vocabulary.
• Run the misread test. Show the mark at 64 px for half a second. If it reads first as a common symbol (a minus, an emoticon, a menu or hamburger icon, a padlock, a play button, a plus), reject it or change it until it doesn't.
• Build the winner in src/brand/Mark.tsx on a 100-unit grid, with gaps of at least 8 units so it survives being scaled down.
• A palette with character, inside the bans. Derive it from the world of the name and the product (a hillside gives pine, moss and slate; a foundry gives charcoal, soot and a glowing red), not from framework defaults. When the name's world is heat or fire, orange is banned, so go red-hot (a deep red glowing on a true charcoal) or white-hot, never generic dev-tool blue: the colour must carry the name. A blue-grey near-black stage is the stock dev-tool dark theme (palette.py notes it), and in a dev tool red UI tags read as errors; spend the red on the brand's own thing.
• Framework blues (the #3B82F6 / #2563EB family) and flat pure greys read as "default".
• Tint the neutrals toward the world, with a cool green-grey or a blue slate; a dark stage can be pine-black or ink-navy rather than #000.
• references/brand-and-color.md §8–10 has worked palettes.
• Set the lockup like a typographer. Render three settings side by side (mark-to-cap ratio, gap, weight, tracking) and pick the one where the mark and the word read as one object. A carelessly set wordmark costs more than any animation gains.
• Render 2–4 one-frame style stills from the Stills folder and look at them: the mark at 16 px and at full size, one statement frame, and one product frame. npm run stills renders the mark stills; add your own with npx remotion still <StillId> out/stills/<name>.png.
• Gate:
• the stills pass §4;
• the mark sheet and the chosen direction's reason are in TREATMENT.md;
• node scripts/lint-film.mjs src reports no placeholder and no colour or font that is not a token.

Step 3: Timeline → src/timeline.ts

Read references/timing-grid.md now.

• Choose BPM and fps so that a beat is a whole number of frames (see §5).
• Map the beat sheet onto bars:
• ACT holds contiguous, bar-aligned acts;
• CUE holds named frames that sit on the grid;
• COPY holds {id, text, in, resolved, out} for every line;
• TOTAL includes a tail of at least 60 f.
• Give each bar one motion peak.
• Gate: npx tsx scripts/grid-check.ts src/timeline.ts passes (npm run check runs it together with tsc and the lint). It checks that every cue is on the 16th grid (or carries an offgrid: reason), that text holds are long enough, that no gap without an event runs past 48 f, and that the word count is within budget.

Step 4: Build, act by act → src/acts/

Before you write motion, read references/motion-tokens.md, references/camera.md, references/transitions.md, and references/product-ui.md if the product appears. Keep references/chromium-rendering.md open while you debug.

• Make each act a pure function of useCurrentFrame(). Mount one <Sequence> per act in Film.tsx, and register each act as its own composition in the Acts folder so you can iterate on it alone.
• Take all motion from the tokens in src/lib/anim.ts (§6). A raw bezier inside an act is a lint error, because that is how a house style decays.
• Build the product UI from real components fed by one data.ts module with asserts on dates, weekdays, counts, plurals and sums. The data must agree with the story too: the person who paid doesn't owe, a "sorted" list is actually sorted, a stated ratio matches its numbers, and no number appears twice in one frame. That holds in every frame, in-between states included: a copy headed "owes Maya" over the whole bill's total is a wrong number on screen for as long as it shows. Frame the UI large enough to read on a phone: push in 2–5× on every interaction that carries the story, and hold a wide shot of readable UI for at most about 1.7 s. Every bar, segment, chart or bare number that carries meaning shows its value, label or unit: a row of "18 24 12 31" under M–S reads as dates, not pages. A screenshot never carries the hero shot.
• Make the product and the feature identifiable. Name the feature once, as a UI label, the one line of copy or the end card, and show enough real app chrome that a stranger knows this is software. "No feature name as a headline" never means "never name it". A feature label is not a payoff: land the value with one short line that makes it human ("Sam had the salad."), even in a muted loop.
• Frame only the meaningful part of the product. No half-empty grids, no tables cropped at the frame edge, no rows of blank cells: they read as a spreadsheet, not a product.
• Text that must be read stays readable in motion. Text being read moves at most about 3 px/f (a camera cruise under it 1.5–3 px/f). Keep labelled chips and numbers under about 20 px/f while they travel, or fly them with the text faded and let it resolve on landing, because motion blur on small text reads as a double image.
• Moving elements never cover text they pass over: plan the paths and the z-order, and check the densest frames at full size.
• A progressive text reveal (letter by letter or with a sweep) must never spell a different word partway through ("tarn" briefly reading "tar"), and never shows part of a glyph: a soft wipe through an "n" reads as an "r" ("halder"). Step the mask from glyph edge to glyph edge on whole frames, or reveal whole words, and check the intermediate frames.
• Export from its act every frame the sound needs, as a named constant (for example LANDINGS or SNAP), so the cue export can import it.
• Wrap everything in <FontGate>.
• Iterate each act in this loop:

1. 1Render stills at each cue −2, 0 and +2.
2. 2Make an act contact sheet: python3 scripts/sheet.py --comp Act2 --every 2 --out out/qa/act2.png renders the act and tiles it with frame and time labels. Use --every 2 for fast sections.
3. 3Run bash scripts/layout-audit.sh Act2 --cues.

• Gate:
• lint-film.mjs is clean;
• layout-audit reports 0 safe-area or overlap violations, and no readable text faster than 20 px/f;
• every act sheet has written review notes.

Step 5: Sound → public/audio/soundtrack.wav

Read references/sound.md now.

• Run npx tsx scripts/export-cues.ts, which writes out/cues.json. Sync points are computed from the picture code (spring contact frames, 50% pop frames, velocity peaks, 97% settles) and never typed by hand. Typed sync points landed 8–11 f off in practice.
• Pick the sound's personality row in references/sound.md §5.5a first. A calm brand gets soft mallets and one gentle chord; drops onto silence and sub booms are for energetic launch films only.
• Make the music drive the picture, whatever the personality.
• A motif that follows the story's progress, for example one tuned note per step climbing the scale (progress in score.json).
• Section changes and the hero reveal land with the music's own event (drop, stop, re-entry). Designed effects go on discrete state changes only, about 5 per 10 s, clustered where the music is sparse; continuous motion and most cuts get none.
• The drop on the hero reveal, 30–60% in, is the loudest moment, and payoff points there. The lockup is the most resolved moment, not the loudest: the music stops on a bar line or filters down as the logo forms, 0.2–0.7 s of near-silence, then one soft resolved element on the frame the wordmark becomes readable (its settleOf, never the tween's last frame or a later bar line: a sound after the motion has stopped lands on nothing) and a 0.5–2.5 s tail. A sting is all lockup: its hit is its loudest moment.
• The sound enters with intent on frame 0 (a drone or a 150–450 ms entry at most), without a click, and runs 5–12 LU under the body for the first 2–3 s. That assumes frame 0 moves: a drone at full level over a still frame (an empty page, a blinking caret) reads as a mistake, so move the picture, not the sound.
• Fill in audio/score.json: key, one chord per bar, sections (with a breakdown or stop about 5 LU down at 40–80% of the runtime), drops and silences. Keep "typing": "auto" unless the brand asks otherwise.
• Run python3 scripts/audio/score.py --cues out/cues.json --score audio/score.json --out public/audio/soundtrack.wav --stems out/stems --json out/qa/score.json (npm run audio). The stems let av-audit.py check each sound against its cue; the master chain lives in scripts/audio/master.py.
• Gate:
• the WAV measures its target ±0.5 LUFS integrated (master.lufs in audio/score.json: −14; −16 only for a sting that plays at the head of other videos, never for a social or feed cut, whatever the brand's temperament), with true peak ≤ −1.5 dBTP before encoding;
• a film of 20 s or more has contrast: LRA 5–8 LU, and the payoff 2–3 LU above the median momentary loudness (score.py warns);
• every sound has something visible that causes it, at a loudness that matches that cause (a caret blink whispers; the hero reveal drops); not every picture event gets a sound, but once a class of action is sounded every instance is, at one level.

Step 6: Render and finish → out/film.mp4

Read references/finishing.md now.

• Preview: bash scripts/render.sh Film out/preview.mp4 --preview --audio public/audio/soundtrack.wav. Remotion renders the picture muted and BT.709-tagged, and ffmpeg muxes the audio. Iterate and critique on previews only.
• Master: if anything moves faster than 12 px/f (true of almost every launch film), bash scripts/render.sh Film out/film.mp4 --blur: sharp render → speed measurement → FilmSub sub-frame render → float accumulation → mux → sync check. Otherwise, with no flag, render.sh renders a sharp master at CRF 14. Loops take --loop (lossless PNG intermediates, so the seam survives the encode) and, when silent, --no-audio.
• Blur costs render time, so spend it once.
• Render the blurred master once, at the end, after the last critique round. It prints the sub-frame multiple and an estimate before it starts; on 4 CPUs a 10 s film at 60 fps and 6× takes about 7 minutes on simple frames and 10–15 on dense UI. --budget 6 caps it.
• A late fix re-renders only the changed act's range (--frames A-B) and splices it, rather than re-blurring the film.
• A 9:16 or 1:1 variant with identical timing reuses the master's samples (--samples-from); one whose fastest motion is ≤ 12 px/f renders sharp.
• Don't edit src/ while a render runs: render.sh renders from a bundle frozen at its start and warns if the source changed meanwhile.
• Stop a render by its own PID, never with a pkill -f pattern: on a shared machine the pattern also kills other people's renders.
• Speed ceiling. Moves over about 80 px/f are listed as too fast for clean blur, and render.sh --blur stops before the sub-frame pass while any are left: redesign them rather than adding samples (§6), or waive a frame-filling edge you checked at full size with --accept-fast A-B. Every hard cut is an act boundary: a cut inside an act blends its two shots into one double-exposed frame, so the render also stops on one (--accept-cut A-B for a flood or flash you checked).
• Gate. render.sh runs both checks below and exits 1 if either fails. Run them again on any file you deliver:
• python3 scripts/probe.py out/film.mp4 --spec 1920x1080@60 --dur <s> passes: size, fps, duration ±1 f, yuv420p, BT.709 tags, AAC at 48 kHz;
• python3 scripts/check-sync.py out/film.mp4 public/audio/soundtrack.wav reports a lag of ≤ 48 samples and true peak ≤ −1 dBTP after decoding.

Step 7: Review loop

Run §8 on previews. Launch films and walkthroughs get at least 2 rounds. For stings and loops whose scripts all pass, 1 round is enough. The blurred master then gets a check, not a round: forensics, av-audit, its fastest frames at full size (every act's, including the lockup's entrance and exit, where a sliding wordmark most often shows stepped copies; fix with measure-speed.py --floor a-b:n), and a director's glance.

Step 8: Deliver

• Run bash scripts/deliver.sh out/film.mp4 with the flags your format needs: --loop --gif --webm for loops, --alpha StingAlpha,StingAlphaClear for ProRes 4444, and --variants Film9x16,Film1x1 for social versions. It writes the poster (the final frame) and, in qa/, a manifest. For a brand you invented, bash scripts/brand-kit.sh exports the mark, lockups, favicon sizes and avatar. Re-lay the social versions out from the same timeline; never crop the master (see references/formats.md).
• Write a short README.md: the idea, the device, the grid, how to render, and the file list.
• Keep the delivery folder clean. Put the deliverables at its top level (the film, alpha versions, poster, logo files) and QA material (sheets, crops, reports) in a qa/ subfolder. A client opens the folder before they open the film.
• Gate: every deliverable in the §1 row exists and passes probe.py at its own spec.

Scaled-down path (stings, loops, short feature clips)

Small jobs should stay fast. The laws, the grid, the lint and one review round still apply. The rest shrinks:

1. 1Write a one-page treatment in TREATMENT.md: logline, device, the pick.py draw for the format, a bar table of 2–8 rows, and the copy. It replaces Steps 1–3 as separate passes. Still write timeline.ts and run grid-check.ts.
2. 2For style frames, the mark at 16 px and one hero still are enough.
3. 3Build the whole piece as one act, or two.
4. 4Sound is optional for a loop. A sting needs one hit tuned to the key, plus a tail that decays to digital zero.
5. 5Render a sharp master unless something passes 12 px/f. Then run one round with the director and forensics lenses.
6. 6For a loop, the last frame must flow into frame 0: the seam's frame difference must be at most max(0.4, 1.5× the median step at the ends) (forensics.py --loop checks it, and so does deliver.sh --loop); render it with render.sh --loop. Build the loop as a cycle rather than as enter, hold and exit. Frame 0 is also the poster and the first frame a visitor sees, so compose it with the product's name or chrome on screen.

4. Taste rules that matter most

The full catalogue of cheap-looking tells, each with its fix, is in references/taste-and-slop.md. Search it during review.

These rules ban the tells of generated work, not personality. Two films made with this skill for different brands should not look alike. If yours looks like the starter, the Tessel example or your last film, it is not finished.

Copy (references/copy-and-type.md)

• Open inside the problem, never on "Introducing…", a rhetorical question or the logo. A muted feed decides in the first second.
• Pay off the film's own words: a setup phrase, then a refrain, then the resolution.
• Avoid stock phrasing (seamless, unlock, AI-powered, streamline) and never use a feature name as a headline. They mark the film as a template.
• Show one line at a time. No eyebrow or kicker labels and no stacked taglines, because that is landing-page grammar.

Type

• Choose one clean, modern sans for the brand's personality, at 2 weights, plus an optional mono, shipped as local variable fonts (@fontsource-variable/*). No serif and no italic (see Hard bans).
• Personality comes from the sans you pick, its weight, case and spacing: a humanist sans at a light weight for calm, a tight grotesk for fast, a rounded sans for friendly.
• Condensed display faces are a generated-look tell.
• The ubiquitous UI default families, the starter's stand-in included, read as "template" unless you chose them on purpose.
• Weight carries tone: a calm brand rarely wants a heavy grotesk wordmark.
• Get emphasis from size or motion.
• Statements are 88–128 px with leading 0.95–1.05.
• Emphasis inside a line is 1.5–1.7× the statement size; a standalone hero word may run 2–5× (14–48% of the frame height); secondary lines are 44–56 px.
• UI must be at least 22 px on screen after camera scale.
• Every number uses tabular figures.
• Keep one tracking token per role, for example a display setting of 600 / −0.05 em. Near-miss values across scenes read as "almost matched".
• Blur text only on entry, and only while it moves fast: statements 6–8 → 0 px, body lines 4–6 px, or no blur at all, through blurIn (clear by 60% of the eased move, because Chromium steps blur radii and renders anything under ~0.75 px fully sharp). Text lands sharp and is never left readable-but-blurred for more than 6 f.

Colour (references/brand-and-color.md)

• Pick one accent from an allowed hue family (red, green, teal, blue, yellow) for a stated reason. Everything in the Hard bans is out: orange, amber, beige and cream, neon, purple, violet or indigo, multi-hue gradients and glassmorphism.
• Restraint is not the same as colourless. The accent visibly carries the key beats (the product's action and the payoff), and the stage and neutrals have a tone from the brand's world. Check a mid-film frame at thumbnail size: if no brand colour shows, the film has no identity.
• Tint the neutrals cool or keep them neutral, never warm or beige, and keep ink as ink. New things arrive in the accent and relax to neutral over 10–30 f, which is how colour says "just happened".
• No glow, halo or bloom. Shadows are 0 20–40px 60–120px rgba(0,0,0,.12–.22), dark surfaces get a 1 px top hairline at 8–12% white, and gradients ship as dithered PNGs (scripts/dither-gradient.py), because CSS gradients band at 8 bits.

Layout

• Avoid "text left, UI card right" and small centred floating cards (under about 40% of the frame width) on an empty field; a centred product window at 70–92% of the frame width, ideally overflowing an edge, is the norm. Show the product full-bleed, anchor the type to an eye line or the lower third, and give each shot one focal action.
• Keep safe margins of at least 96 px at the sides and 64 px top and bottom, and at least 24 px of headroom at maximum punch. Overscan moving layers by 5–8%.
• Fill the frame with purpose.
• In product shots the hero (the UI, the object, the number) covers about 40–70% of the frame.
• Outside designed negative space around a single hero, such as a logo, no region of more than about a quarter of the frame stays empty.
• In vertical feeds, the platform's caption and button zones get background or continuation (the product's surface, the stage, a bleed), not an empty band. Viewers read emptiness as unfinished, whatever the reason.

Decoration

• Every element must name its job (reveal, route, validate, emphasise) or be cut. That rules out particles, bokeh, starfields and ghost text.
• Grain is either absent or global at 1.5–2%, keyed to the output frame.

Product truth (references/product-ui.md): named, specific, consistent data; counters count on each visible event; the product only replans the future; no cursor once the product acts on its own; any paused frame still makes sense.

Transitions (references/transitions.md)

• Use 6–8 types per film, each at most twice, plus one signature move taken from the mark and used 3 times (open, middle, close).
• A crossfade is never the default, and never goes through black (it dips about 25% in luminance).
• No preset gimmicks: glitch, light leak, film burn, swirl, ripple, dreamy zoom, flash through white.

5. Timing system essentials

src/timeline.ts is the only source of timing. Picture, cue export and grid check all import it. Details and the full pattern are in references/timing-grid.md.

export const BEAT = (60 / BPM) * FPS;               // 30 f at 120 BPM / 60 fps
export const BAR = BEAT * 4;
export const b = (bar: number, beat = 0, sub = 0) =>   // sub = 16ths
  Math.round(((bar - 1) * 4 + beat) * BEAT + sub * (BEAT / 4));
const L = (abs: number) => abs - ACT.plan.from;         // act-local frames, one per act

• Whole-frame beats at 60 fps: 90 BPM gives 40 f, 100 → 36, 120 → 30, 144 → 25, 150 → 24.
• Grid placement
• Section changes land on downbeats; cuts inside a section land on 8ths.
• A batch of landings snaps to 16ths with a 0–3 f spread, and the first item lands on the beat.
• Micro-rhythm is welcome: the payoff on the clap, its consequence on the next kick.
• Cuts land on the grid, but most carry no sound of their own: the music carries them. Section changes and the hero reveal land with the music's own event (drop, stop, re-entry); av-audit.py suggests one when a top-3 picture change has no audio onset.
• Hit frames
• Start a picture change at cue − 1, because prog() is 0 on its start frame.
• Pulses peak 2 f after their sound (hitPulse).
• Start a spring at beat − delayTo(cfg, 0.5) so its pop lands on the beat.
• Holds
• Once text resolves, it stays at least 36 f + 6 f per word. Text that hasn't resolved never exits.
• Nothing is static for more than 48 f, except a designed freeze of ≤ 15 f that sits exactly on a silence, and one designed near-still hold of 60–96 f per film on the key claim or success state, under a music dropout and followed by a big move (declare it as export const HOLD = {from, to} in timeline.ts). When a gate flags a quiet stretch, give it motion; loosening the threshold to pass is not a fix.
• The end card runs 1.5–2.5 s: the lockup assembles over 60–80 f, then holds still for 36–80 f or keeps building (a creep of ≤ 6%, or a 10–20% push when the hold runs past about 1.3 s), then the tail. Never cut to a complete card that sits static, and never more than about 2.5 s on a still end card: a longer one reads as a drag. A URL is readable for at least 1.7 s.
• In a launch film or product video the lockup appears once, at the end (a sting is the exception: it is all lockup). A mid-film logo reveal followed by the same lockup again reads as a repeat and spends seconds on a still; the one exception is the product's name or mark revealed at about 20–26% of the runtime after a problem act of 10 s or more, when the end card is a different construction or mirrors the opening. Mid-film, the brand is present through the device and the accent. Brand logo time across the whole film stays ≤ about 12% of the runtime.
• Density
• Aim for 45–60 discrete events per 30 s and one motion peak per bar, with 30–70 f of calm between peaks.
• The first 6 s run about twice as dense as the middle.
• Energy builds toward the payoff. No stretch after the turn is slower or greyer than the one before it, except one designed lull of 1–4 s inside a music breakdown at about 55–85% of the runtime; the last and biggest visual peak lands at 75–90%, on the music's re-entry. A second half that drags loses films whose first half is strong. Check the energy by eye on a contact sheet of the second half.
• The last third keeps moving: a camera push, secondary motion or the device still acting. A film that settles 4 s before the end has ended early.
• Shot length: median 60–90 f; minimum 24 f (and only with no text); maximum 180 f in a cut-driven film, while a continuous take may run 300–560 f if an in-shot beat lands at least every 72 f.
• Hook:
• frame 0 is already composed and moving;
• the sound enters with intent on frame 0, with no fade over the first bar;
• something visibly changes about every 0.5 s;
• the problem is felt by 1 s and stated by 2 s;
• the product is named on screen by about a quarter of the runtime (its chrome, the name in a status line, the mark), not saved for the end card.

A soft first second loses a muted feed.

• End card: carry the brief's one practical fact when it gives one (a date such as "next week", a URL, "out now"), set at the secondary size under the lockup. When the brief gives none, a launch film or product video still says what the product is in one short line under the name ("The build cache for CI."); a bare name is for stings. That is the only line besides the name.
• State follows the event: counters, checks and totals change on the landing frame of the thing that caused them, not a beat later.

Where a sound goes (computed by src/lib/sync.ts: peak, hit, delayTo, settleOf):

Audio that leads the picture by more than about 45 ms reads as "sound first". Move the picture, not the sound.

6. Motion system essentials

The tokens live in src/lib/anim.ts (E, SPR, tw, prog, mix, lmix, hitPulse, blurIn, arrive, arriveK, stagger, inertia, rand, fd), and tested primitives for the commonest moves in src/fx/ (Words, TypeOn, Scramble, Roll, Morph, Iris, MaskRise, Cursor, RackFocus): use them before writing your own. Curves, measured values and choreography constants are in references/motion-tokens.md; camera work is in references/camera.md.

A film uses about 3 easing characters. Linear is only for drift.

Springs (SPR, as damping/stiffness/mass): snap 18/260/0.7 (small overshoot), pop 11/180/0.6 (a landing only), soft 26/120/1 (no overshoot), heavy 30/90/1.4 (weighty).

• Damping. Everything is eased or critically damped: damping ≥ 2√(stiffness·mass). Overshoot (1–7%) belongs only to a true landing, an object arriving at a surface or a lock, at most 2 per film, because a bounce anywhere else reads as a toy.
• Smoothness is a ship gate. Jitter, pixel-snap stairs, single-frame pops and stall-then-lurch block the ship however good the idea is: zero forensics.py fails and Finish at 5 (§8).
• Clamp. Clamp overshoot wherever it could collide with other geometry or drive a value negative. Snap to rest when |1 − s| < 0.01.
• Easing.spring({damping:200}) starts at zero velocity, so it is not a replacement for expo-out.

Laws of weight:

• Direction. Entrances decay and exits accelerate; only the camera eases both ways. A fade-out never uses expo-out, because it pops half the change in one frame.
• Log space. Interpolate scale and zoom in log space: exp(mix(log a, log b, t)), which is lmix.
• Duration. Duration grows with distance: 0.35 s + 1.35 ms per px, clamped to 0.6–2.4 s for camera moves. The slowest move is at least 3× the fastest.
• Shape and contact. Shape settles 6 f before position. Contacts squash (scaleX 1.08 → 1, scaleY 0.9 → 1 over 8 f, anchored at the contact edge).
• Overlap. A secondary move starts at 55–75% of the leading move (about 20 f into a 60 f move), so both settle within 2 f, or it starts on a zero-slope curve. This prevents stall-then-lurch.
• Budget. Run one hero motion plus at most 2 supporting ones. Put fast (60–70 px/f whips) next to real stillness (0.3–0.6 px/f). Anything faster than 12 px/f gets true motion blur in Step 6.
• Ceiling. No element moves faster than about 60–80 px/f. Past that, blur smears it into a streak with stepped copies; redesign the move as a cut on the beat, a match cut, a mask wipe or a shorter distance. Frame-filling edges (wipes, floods, irises, zoom-throughs, a container growing to full frame) may peak at 150–300 px/f for a few frames.
• Envelope. Every punch and tick uses hitPulse(t, attack = 2, tau = 5). A half-sine on a hard window leaves velocity clunks and peaks 6–8 f late.
• Seams
• Either both sides of a cut are at rest, or the incoming shot starts at the outgoing velocity.
• Match every property across a cut, and show the rest pose once.
• Never cross-fade two copies of one object; morph one object instead.

7. Engine non-negotiables

These come from real failures. Each costs minutes to respect and hours to debug.

Both engines

• Everything is a function of frame or time.
• No CSS transition, animation or @keyframes, because they don't render deterministically.
• No Math.random, Date.now or performance.now; use a seeded rand(seed).
• Transforms
• Put all translation inside one transform (left:0; top:0; transform-origin:0 0; transform: translate() scale()). A fractional left/top combined with a transform snaps to whole pixels.
• Place dots and cursors by transform only.
• Vertical text positions snap to whole pixels in the headless shell, so end a text rise on arrive(p) (and start an exit on depart(q)) instead of letting an ease-out creep its last pixels (references/chromium-rendering.md §15).
• Blur
• Never put will-change on a layer where filter: blur() animates. It causes nondeterministic ghost frames.
• Never animate blur on large layers; use src/fx/RackFocus.tsx.
• Discrete state
• Clamp every value that feeds geometry, so radius, width, blur and scale stay ≥ 0.
• Compute discrete state (typed count, caret, counters, labels) from fd(frame) (Math.round), because motion-blur sub-frames are fractional.
• Fonts. Load fonts before any render or text measurement (FontGate). A measurement taken earlier is cached with the fallback width.

Remotion (references/remotion-engine.md)

• Clamping. interpolate extrapolates by default, so always clamp; tw() and prog() do it for you.
• premountFor only affects Studio preview and does nothing in a render. With layout="none" it is silently ignored in 4.0.529 (no error). Don't build on it.
• *Studio `Interactive.` markup** is an optional leaf layer, never the structure.
• Config. Render with --muted and let ffmpeg mux the audio (Remotion's AAC leaves about 2.5 f of priming). Use the ANGLE GL renderer, JPEG q95 intermediates and a 120 s delayRender timeout; the starter's remotion.config.ts sets all of these.
• Browser. In a sandbox, set REMOTION_BROWSER to the local headless Chromium if auto-detection misses it.

HyperFrames (references/hyperframes-engine.md)

• Timeline. Each composition has one paused timeline, registered last, with a key equal to its data-composition-id.
• Tweens
• Use fromTo, never from.
• Never combine a CSS transform with a GSAP transform on one node.
• Markup
• No repeat:-1. No <br>.
• Every <audio> has an id; one without is silently dropped from the mix.
• A sub-composition's <style> and <script> go inside its <template>.
• Fonts are local woff2 files declared with @font-face.
• Lint. Lint must show zero errors before check means anything, because otherwise it reports a misleadingly clean "0 samples".
• Finish. Finish with bash scripts/hf-finish.sh. Films faster than about 18 px/f belong in Remotion.

8. Review loop and ship gate

The full procedure, the critic prompts, the rubric anchors and the table of thresholds are in references/review-loop.md. A round goes like this:

1. 1Render. Render the preview (or the affected acts).
2. 2Measure. Run the scripts. Each writes JSON to out/qa/: - python3 scripts/probe.py out/preview.mp4 --spec 1920x1080@60 --dur <s> - python3 scripts/forensics.py out/preview.mp4 --cues out/cues.json --json out/qa/forensics.json. It checks pops, stalls, dead holds, ghost frames, border slivers, judder, sharpness steps, banding and compression smear on flat fields. Declare designed cuts and freezes with --cuts and --freeze-ok so they are not flagged, and add --loop for loops to check the seam. - python3 scripts/av-audit.py out/preview.mp4 --cues out/cues.json --stems out/stems --json out/qa/av.json. It checks onsets against cues (per stem, so the bed cannot hide a late effect), loudness, true peak, clicks and the tail.
3. 3Sheets. Make contact sheets with python3 scripts/sheet.py out/preview.mp4 --chunks 4 --rate 12 --out out/qa/sheet.png, plus a legibility sheet at --width 480. Grab exact frames with bash scripts/grab.sh out/preview.mp4 <frame>….
4. 4Lenses. Run the critic lenses from assets/critics/. Launch them as parallel subagents if the harness allows it; otherwise run them one after another, and look at the sheet before you open the code. On a tight usage budget, run the director and art lenses only, yourself, on the sheets, and let the scripts cover forensics and sync; never skip the lenses entirely, because they catch most of what the scripts can't. - Director (director.md): the weakest 3 s; anything cheap, templated or dead; whether each act lands in under 1 s; whether the story reads muted; the 2 s hook; the end. At most 12 changes. - Forensics (forensics.md): reads the script JSON and inspects the flagged frames at full size. - Sound/sync (sound-sync.md): onsets against cues, masking, key, the ending. - Round 1 only, art/copy/UI (art-copy-ui.md): type, colour discipline, strings, data truth.
5. 5Output format. Each lens returns {issues:[{id, priority, frame, problem, fix:{file, change}}], overall, score}.
6. 6Verify (verifier.md). Try to refute each P0 and P1 from frames and code. Drop an issue if it is wrong, invisible at normal speed, or if its fix would make things worse.
7. 7Fix and re-render. Fix in order of impact per minute. Re-render only the affected acts (per-act compositions) and splice at a static seam frame.

Critic prompts say what changed, so it gets checked hardest. They also say "Do not re-report fixed issues" and "Don't propose adding text", and they batch images into sheets, because images are expensive in context. When one act stays weak after two rounds, run the design-off in review-loop.md.

Severity: P0 is visible on a key beat (hook, payoff, logo) or breaks the story; P1 is noticeable on a normal viewing; P2 is visible only when paused.

Ship gate

• The 11-dimension rubric: every dimension ≥ 3, Finish and Sync = 5, mean ≥ 4.2.
• No open P0.
• Every script passes.
• Stop after 5 rounds, or sooner once the gate is met. Then say plainly what you would still change.

9. File index

Read each file when its step comes up. Don't read them all at once.

Tools you already have. Reach for these before writing a helper of your own.