alifmahmudashik/shorts-studio-skill

Cartoon 3D story short (three.js + Remotion)

Proven example: The Bridge That Fell for a Song. It is an original 67 s short about the 1831 Broughton bridge and 74 soldiers marching in step.

• Code: src/videos/bridge/ of your Remotion project (composition BridgeShort, 1080x1920, 30 fps).
• Script, voice and sound: ~/bridge-short/.
• Style research: references/.

The style reference is a channel studied privately (references/: STYLE_BIBLE.md, reports/*.md, frames/).

• Those videos are private study material. Never upload, publish or send them.
• Never copy their scripts, characters or shots. Story, characters and voice are always OUR OWN.
• Never say the channel's name in a voice-over or description.

User rules that apply:

• Captions: ON by default for Shorts. One word at a time, centred, bold, dark outline, inside the safe area, timed to words.json (see 'Subtitles'). Only leave them off if the user says so. No other on-screen labels.
• English only.
• Voice: one steady tone (deadpan: the narrator never reacts, the characters do).
• Jokes are researched real facts.
• Standard mix: music +3 dB, SFX -6 dB, voice unchanged.
• Final export 2K vertical: 1440x2560, --scale 1.3333.
• Opus max verifies every check.
• Save after every step.
• "Just do it" once the script is approved.

Pipeline

1. Topic and facts. Pick one surprising true story with a physical cause and effect (it should be visual), plus a simple everyday analogy that explains it.
   • Write the facts with their source at the top of SCRIPT.md.
2. Script (~3.7 words/s, 35-70 s). For fact shorts use 'Viral story format' and 'Make it funny' below; the bridge beat formula is the alternative for disaster stories:
   • Hook: the absurd fact in one line ("Seventy-four soldiers knocked down an iron bridge just by walking across it.").
   • Setting: year, place, pride.
   • Action: escalation over three or four beats, each a little bigger.
   • Snap: the moment it breaks.
   • Deadpan double line: "Luckily ... / Unluckily ...".
   • Explanation: one physics idea through an everyday analogy (the kid on a swing).
   • Consequence: the rule it created.
   • Callback twist: the ending.
   • Never describe on screen what the viewer already sees.
3. Voice. Record or generate voice.wav, then make its word timings words.json (any speech-to-text with word times).
   • Put them in the project folder, e.g. ~/bridge-short/voice/.
   • Hand-fix the word timings. Everything is timed to words.json, and the captions are drawn from it too.
4. Shot plan in story.ts: copy templates/bridge/ (the final bridge source, same as src/videos/bridge/ of your Remotion project) to src/videos/<name>/, register a new composition next to BridgeShort in src/Root.tsx, and rename it.
   • SHOTS: id, a/b frames, light day/golden, crowd mode, tau story time.
   • XFADE: crossfades at place/time changes.
   • FLASH_AT: a white flash after the hook.
   • CAMS: per-shot keys {f, dur?, cut?, s: f => {pos, look, fov, roll}}.
   • KICKS: impact flip-shakes.
   • DOF/dofAt: lens blur per beat.
   • FOREGROUND: near framing props.
   • Every visual reaction lands 0-3 frames before its word.
   • Chain the subjects of each sentence into one continuous flow (flow.ts, see 'Continuous flow'), give every move a name and every stop a framing size (see 'Camera variety' and 'Framing').
   • Give every place its own light preset (see 'Lighting for every place').
5. World.
   • lights.ts: LIGHTS plus the LOOK sky/river/cloud palette per light.
   • shade.ts: two-tone shading and untone().
   • build.ts: town, houses and Kenney CC0 kit props, recoloured, with sight-line culling of anything blocking a camera.
   • paint.ts: HD Canvas2D painted textures with seams painted in as dark tints.
   • world.ts: sky shader, fog = horizon colour, river, per-frame posing, sun frozen per shot and camera-relative.
6. Characters.
   • toon.ts: the Canvas2D Pen drawing API, 600x1000 design canvas, soles at y = 975, bean heads, stick limbs, plum variable-width outline, face library.
   • cast.ts: camera-facing billboards, crowds, effect cards, blinks, walk cycles. All billboards go on the NOINK layer.
   • cardshade.ts: cards lit by the scene light, no emissive.
7. Post (Bridge.tsx + ink.ts + lensblur.ts + grade.ts).
   • Composer order: ink → lens pre → lens main → bloom → NEUTRAL tone map → grade (MSAA 4).
   • Every pixel parameter scales with output height / 1920.
8. Sound (~/bridge-short/audio/: timeline.py, sfx_gen.py, sfx.py, music.py, instruments.py, dsp.py, mix.py).
   • Event table → synthesized SFX plus an original score ducked under the voice.
   • Output: mix.wav (48 kHz, exact length), stems and cues.md.
   • Lock the picture's time constants before the sound is made. Any later timing change means re-running the sound.
9. Look checks. Run node scripts/bridge-stills.mjs out/<name>/x_ 100,400 0.5.
   • Set BRIDGE_ENTRY=src/sandboxes/<n>/index.tsx to render a sandbox copy.
   • Make side-by-sides against reference frames with python scripts/sbs_sheet.py out.jpg --ours ... --ref ....
   • Send them to an Opus max verifier: 1-3 images, a ≤200-word verdict.
   • Run the framing check (templates/bridge/framing.ts) for every shot and make one contact sheet with a still per shot (first, middle, last frame). See 'Quality gate'.
10. Final render.
    • npx remotion render src/index.ts <Comp> out/<name>/video.mp4 --codec h264 --crf 16 --scale 1.3333
    • Mux with -map 0:v:0 -map 1:a:0 (Remotion writes a silent audio track).
    • QA: python scripts/short_qa.py final.mp4 out/<name>/qa shots.json checks every frame for black, white, frozen and pop frames, plus loudness and sheets.
    • Fix anything it finds, then run a deep Opus audit.
11. Deliver. SendUserFile the final mp4, update memory, write notes.

Preflight (do this first)

• Check that templates/bridge/, references/ and scripts/ exist (clone github.com/mromidev/claude-animated-shorts, or your fork, next to the Remotion project if they are missing). If they are missing, STOP and fetch them. Never improvise a generic three.js look without the templates.
• Make a 15 s test clip first (hook + 2-3 shots, with captions and sound). Render stills, run the Quality gate, fix, then extend to the full story. Never go straight to a full render.

Viral story format: "laugh at a genius" (use for fact shorts)

Six beats, every time:

1. Weird action with no context, shown in one picture (hook, first 3 s).
2. The laugh: the other characters mock the hero.
3. They all look right: time passes and the hero seems wrong or harmless.
4. The clue: one detail that foreshadows the payoff.
5. The payoff: the event that proves the hero right (the visual climax).
6. The science: show the mechanism (inside / cutaway / macro view) and explain it in one simple sentence.

• Tell it as a story with a hero, not a list of facts. Pick a weird practice that is visible in one image and that people would laugh at.
• The last line starts an argument or a comment (something debatable), not a neat summary.
• Title pattern: "Why X ...", "How X ...", "What the X really meant". No spoilers of the payoff in the title.
• Pacing: a new subject every 2-3 s, reached by camera travel (see 'Continuous flow') rather than a cut, with a sound effect on almost every move or dissolve (whoosh, clink, crowd laugh) and music under the voice.
• Rich places, simple characters (round heads, stick arms): plan 12-16 distinct places and 20+ props for a 60 s short, and give every place its own lighting mood.
• Write the facts with sources first. Years and numbers are spoken naturally ("eighteen forty-seven") and captioned as digits.
• Example: templates/SCRIPT_viral_example.md.

Make it funny (comedy engine)

The crowd laughs at the hero; the viewer should laugh too.

• Deadpan voice, absurd pictures. The narrator never reacts; the characters do. The gap between the two is the joke.
• A gag every 4-6 s, using at least three kinds per short: a line gag (understatement, "Luckily ... / Unluckily ..."), a visual gag (reaction face, a tiny detail in the background, a wrong-scale prop) and a sound gag (a silly SFX that lands on the reaction).
• Rule of three: two normal beats, then the third breaks the pattern.
• Running gag: one background character or object (a pigeon, a man with a sandwich) shows up 3-4 times and gets a payoff near the end.
• Reaction shots: the camera settles on the face (a flow stop or a rack focus, not a cut) 3-6 frames after the punch word and holds 8-12 frames. Use the face library; include at least one "unimpressed" and one "shock". See 'Reactions and dialogue beats'.
• Confident hero, smug crowd, over-serious officials: characters take trivial things deadly seriously.
• Exaggerate physics for one beat only (the payoff): squash and stretch, a dust cloud, an impact kick.
• Callback ending: the last line twists the first line (the hook) using what happened.
• Jokes stay inside real facts: the funny part is the true detail. Never punch at real victims or at groups; the target is overconfidence and smugness.

Subtitles (word-by-word captions)

• Source: words.json (voice word timings, hand-fixed). Any speech-to-text works; normalizeWords() in templates/bridge/Captions.tsx accepts {word|text, start|start_time, end|end_time} in seconds or milliseconds and skips ElevenLabs "spacing" entries.
• If the speech-to-text text differs from the script (numbers, names, spelling), replace the text with the SCRIPT words and keep the STT times. Captions show the script spelling ("1847"), even though the voice says "eighteen forty-seven".
• Component: copy templates/bridge/Captions.tsx to src/videos/<name>/ and render <Captions words={WORDS} /> as the LAST child of the composition, as a DOM overlay above the three.js canvas, outside the post-processing composer (no ink, blur, bloom or grade on text).
• Style: one word at a time, horizontally centred at about 68 % of the frame height (below faces, inside the safe area), heavy sans, white fill, 9 px dark outline at 1080 wide, pop-in scale 0.8 → 1.08 → 1 over 6 frames, no background box, max width 86 %. Numbers, names and punch words use the accent colour (highlight). Pixel sizes scale with output height / 1920.
• Never cover a face: when a close-up has the face at the caption position, pass yPercent={(frame) => ...} to move it for that shot.
• The caption appears 1 frame before its word is heard (lead).
• QA: render a still on a punch word in every shot and confirm the caption is inside the safe area and does not cover a face.

Framing, safe area and zoom limits (enforced)

• 9:16 frame. Safe area: 6 % margin left/right, 8 % top, 15 % bottom (platform UI and captions). Everything that matters (characters, props, bubbles, signs) is fully inside it, except deliberate foreground framing props.
• A character's head takes at most 30 % of the frame height (a one-shot reaction gag may use 40 %). Prefer medium shots (waist up) and full-body wides.
• Never cut a character off at the feet, and never crop a speech bubble or sign. A foreground prop is either recognisable or gone; a cut-off prop must not cover more than 35 % of the frame.
• No more than two shots of the same size in a row (wide / medium / close / insert).
• Push-ins on holds stay slow: 0.15-0.35 % per frame, 0.6 % on punch words. No sudden big zooms.
• Automatic check: checkShot(camera, items) in templates/bridge/framing.ts projects bounding boxes to screen space and lists every problem. Run it on the first, middle and last frame of every shot, in bridge-stills, and fix failing shots before rendering. Self-test: npx tsx templates/scripts/framing_selftest.ts.

Lighting for every place (the bridge only defines day and golden)

Every place needs its own light preset in lights.ts: a clear key direction, two-tone shading (warm lit, cool shade), cast and contact shadows, and a matching sky or ceiling. Never a flat unlit room, and never sky or clouds visible inside a room.

• Interior (hall, hotel, office): a ceiling or dark wall top, one key light from a window or lamp, warm practical lights (candles, lamps) with bloom, a cool shade side, a soft vignette, floor and wall shadows from every prop, light shafts or dust from windows.
• Night: moonlit blue key (about #5b7fd6), deep blue-violet shade, warm lit windows and lamps as contrast, moon disc or stars.
• Sunrise / sunset: low sun with god rays, peach to pink sky, violet shade, long shadows.
• These are starting values: tune them against references/frames/ with side-by-side sheets (sbs_sheet.py).
• Contact shadow under every character and prop, plus a directional shadow tail from the key light; furniture and props cast shadows on floors and walls (card shadows are real, not blobs, in interiors).
• Depth of field (table below) is applied consistently: close-ups and medium shots are never fully sharp, wides stay sharp.

Camera variety (always moving, never the same move twice in a row)

• Within a flow or shot, use named moves: slow push-in, pull-back reveal, orbit around the subject, lateral track with parallax, crane up/down, tilt, rack focus from a foreground prop to the subject, small dutch roll (4-5°).
• No two consecutive moves are the same.
• Put foreground props (pillars, rails, furniture) between the lens and the subject in about a third of the shots, so camera motion reads as depth.
• At least two insert shots per story: a macro or cutaway of the mechanism for beats 4-6, plus one dramatic angle (low hero angle or bird's eye).
• Keep the existing smootherstep glides (18-40 f), creeping holds and impact kicks.

Continuous flow: travel between subjects, don't cut (the reference's signature move)

Measured on two reference shorts of 65 s each: automatic cut detection finds about one hard cut per minute. Everything else is the camera travelling through ONE world from subject to subject, plus short dissolves (3-8 frames) when the place or time changes, and travels through windows and doors that hide the change. A weaker version of this skill cut to a separate shot for every subject (tower, then the people, then a face), which looks cheap.

• Same place, no cut. "Paris was broke": start wide on the Eiffel Tower, then travel down (tilt, dolly) to the crowd at its foot, then settle on one worried face, in one unbroken move while the voice says the line. Tower → crowd → face is a chain of three stops in one flow.
• Build the set for it. Subjects that share a sentence live in the same 3D space: the crowd stands at the tower's foot and the hero is among them. Do not build a separate set for each shot.
• Settle, then go. Each stop is a framing of the subject (wide / medium / close, see Framing) held for 20+ frames with a creeping push-in (0.15-0.35 %/f, push). Travel between stops takes 18-60 frames (checkFlow warns outside 18-75), on a curved path with ease in and out.
• Timing to the voice. The camera arrives on the subject 0-3 frames before the word that names it. Travel happens on the connecting words or in the pause; stops land on the nouns and punch words.
• Rack focus follows. The focus point travels with the camera (flowFocus), sharp on the new subject 4 frames before the camera settles.
• Scale jumps are travel too. Pull from a macro (a hand, a coin, a dragon's mouth) out to the hero and the town, or push from the town into one object that becomes the next scene.
• Place or time change: a 3-8 frame dissolve (XFADE), or hide the change in the move itself: travel through a train window into the car, tilt up into the sky and back down into the next place, pass behind a wall or a prop.
• Hard cuts are only for a gag or an impact, about once per 30-60 s.
• Budget: 4-6 flows per 60 s story, each 4-10 s long and chaining 2-4 subjects; at least 60 % of the screen time is inside a flow.
• How: templates/bridge/flow.ts. A flow is ONE key in CAMS ({f: start, cut: true, s: flow(stops)}), flowFocus(stops) feeds DOF, lookFrom(target, dist, azimuth, elevation) places a camera around a subject, checkFlow(stops) lists problems. Self-test: npx tsx templates/scripts/flow_selftest.ts. Run the framing check (framing.ts) on every stop.

const TOWER = lookFrom([0, 150, 0], 160, 20, 8);   // wide of the tower
const CROWD = lookFrom([6, 1.4, 12], 5, -25, 4);   // the broke people at its foot
const HERO  = lookFrom([8, 1.5, 12.5], 2.2, -10, 2);
const stops = [
  {f: 0,   ...TOWER, hold: 20, focus: [0, 150, 0]},
  {f: 58,  ...CROWD, hold: 24, focus: [6, 1.2, 12]},
  {f: 118, ...HERO,  hold: 30, focus: [8, 1.6, 12.5]},
];
CAMS[3] = [{f: 0, cut: true, s: flow(stops)}];

Reactions and dialogue beats (light touch, no lip-sync, one narrator voice)

Characters do not talk with moving lips and there is no second voice. The narrator reads the line, and the characters sell it with their faces. This uses the existing face library in toon.ts (neutral, stern, grin, laugh, whistle, nervous, shock, relief, dazed, pained, smile, lick); do not add new drawing code unless a reaction is truly missing (then add one Expr entry to the FACE table).

• Reaction faces (mouth and eyes together):
  • laugh → laugh + crowd-laugh SFX
  • shock / gasp → shock + gasp or record-scratch
  • nervous → nervous + gulp
  • dazed after an impact → dazed + ding
  • hurt → pained
  • smug → smile or grin
  • guilty or innocent → whistle
  • unimpressed → stern or neutral, held 12 f
• Timing: the expression changes 0-3 frames before the word it reacts to, using a blink to hide the swap (as in cast.ts). The reaction holds 8-12 frames, then eases to the next expression. Never stay on one face for a whole line.
• Dialogue beat (optional, at most 2 per short, one line of 8 words or fewer each):
  1. The narrator reads the quoted line in the same deadpan voice: "Is this legal?" the dealer asked. "Very," said the con man. Captions use the same style.
  2. The speaker is a flow stop: settle on them while the line is read (20+ frames, creeping push-in).
  3. A silent beat of 6-10 frames with a creep, no new sound.
  4. The camera travels or rack-focuses to the listener (flow stop or flowFocus), and their face lands 3-6 frames after the last word.
  5. Hold the reaction 8-12 frames with a small SFX, then continue the flow.
• Camera for dialogue stays inside the flow rules: speaker and listener are two stops (a close two-shot, an over-the-shoulder, or an orbit between them), never a pair of separate cut shots. The only cut allowed is one smash cut to a reaction for the biggest gag of the short, and it counts toward the hard-cut budget (about one per 40 s).
• Extras that fit the same rules: double take (face looks away, then back with shock), punch-in on the last word (0.6 %/f for 35 f), freeze for 6 f with a record-scratch, a bystander's face in the background reacting a beat late.
• If real lip movement or a second voice is wanted later, add it as its own feature; the default is narrator-only.

Quality gate (before every full render)

1. Framing check passes for every shot and every flow stop (framing.ts). Hard cuts are at most 2 per 60 s, every subject change inside one place is a flow (checkFlow has no warnings), and the flows cover at least 60 % of the screen time.
2. One contact sheet with a still per shot (first, middle, last frame), side by side with references/frames/.
3. Score lighting, shadows, blur, framing, zoom, camera variety, captions and gags out of 10. Fix anything below 8 and re-render those stills.
4. Every gag has a reaction face that changes 0-3 f before its word. Voice 20 dB above music; numbers spoken naturally; captions match the script; a sound on almost every cut.
5. Only then run the 2K render and short_qa.py.

Final look values (bridge after 4 verifier rounds; all in templates/bridge/)

Start every new short from templates/bridge/ (copy to src/videos/<name>/): the look lives in these values, the story lives in story.ts / build.ts layout / cast.ts roles. Opus max scores at the end: colour day 7, golden 5-6, lighting 5-6, ink 6, textures 5-6, characters 7, depth 6-7, composition 6-7 (from 5/3/4/5/4/7/4/6 at the start).

• Light presets (lights.ts, sun camera-relative and frozen per shot):
  • day: az 75 / el 48, lit [1.58,1.5,1.33], shade [.30,.29,.36], lift [.05,.06,.11] (cool tint ONLY in shade), sky top #3e7fd2 → horizon #b8d0ec, fog = horizon, 25→300 m, uSat 1.08, ~15 % blue haze past 60 m.
  • golden: az 6 / el 7 (backlit low sun), key toward #FFB36B ×1.5, violet-blue shade (facade ≈ #444380), sky #ffc47c → #8a6f8a, peach fog only past 50 m, sun disc #ffcf9e + bloom, contact blobs + real card shadows.
• Albedos (paint.ts/build.ts): muted brick / cream / ochre / terracotta houses with plinth, cornice, sills, hoods, door step, chimney + pot, heights 7 / 8.4 / 9.8 m; warm grey/beige stone (saturation −35 % from "storybook"); ochre wood deck C* ~51-58 with near-black seams; iron dark and desaturated. Never lemon, never lavender lit faces.
• Ink (ink.ts): #030303, 5.1 px near (6 m) → 3.3 px far (90 m) at 1080x1920, creases ×0.8 from 35°, depth jumps from 1 %, fog tint 0.3 only far; character cards on layer NOINK (they carry their own plum ink).
• Lens (lensblur.ts + dofAt in story.ts): cInf px at 1080: hook/river CU 15-16, close-ups 12-14, MS 10-12, park 7, golden sway 5, wides/aerials 0. Focus on the eyes of the front row.
• Post (Bridge.tsx): ink → lens pre → lens main → bloom (thr 1.0, int .5, 6 levels, radius .6) → NEUTRAL tone map → grade; MSAA 4. Film grain exists but is OFF (reference measures no grain; it quadruples the file size).
• Characters (toon.ts/cast.ts/cardshade.ts): bean heads h/w .94, plum ink #241726, eyes .26 W apart, no blush, 14 expressions; adults 3.5 head-heights; walk bob 5 %, lean 1-7°, planted 0.8 m stride; blinks 0.5-2.3 s; tight contact shadow ~.5 under the feet + sun-direction tail by day; golden rim inner only (never an offset outer copy).
• Camera (story.ts): continuous flows between subjects (flow.ts, see 'Continuous flow') instead of cuts, short dissolves at place changes, hard cuts only for a gag or impact, smootherstep glides 18-40 f, every hold creeps (no static hold ≥4 f), flip kicks decaying over 7 f on impacts, 4-4.6° rolls, foreground props (FOREGROUND) in ~1/3 of shots.

Look targets (measured from the reference; see references/reports/)

• Ink:
  • World lines: about 3 px at 1080 wide, near-black #0b0a0c, nearly constant with distance. Silhouettes, depth jumps and normal creases only. No line on wall/floor joins or the horizon. Character cards are never outlined by the world ink.
  • Character ink: dark plum #201628-#2f1d28, thicker on the bottom/shadow side.
• Colour:
  • Day median luma ~.57. Non-ink p5 ~.23 (no crushed blacks). Lit chroma C* p90 ~55. Day sky top ~#5e93d9, horizon 1.28x brighter.
  • Golden: peach horizon #ffd08a, violet-blue shadows (never brown), backlit sun.
• Depth:
  • DOF on about 1/3 of screen time (CU, MS, inserts), cInf 10-16 px at 1080, focus on the eyes. Establishing and wide shots stay sharp.
  • Fog = horizon colour, 25→420 m day, 20→360 m golden.
  • Bloom only on HDR emissives: threshold 1.0.
• Camera:
  • Subject changes inside one place are continuous flows (flow.ts); place or time changes are a 3-8 f dissolve (XFADE) or a travel through a window, door or sky; a hard cut is rare (about one per 40 s). Small reframes are a smootherstep glide of 18-40 f, peak ≤ 25 px/f at 720p.
  • Every hold pushes in 0.15-0.35 %/f, with 0.6 %/f on punch words.
  • Rolls of 4-5°.
  • No hand-held float. Impact shakes flip sign every frame and decay over 7 f.
• Characters:
  • About 3.5 head-heights tall, legs 1.4 HH.
  • Stick limbs about 0.055 of head width, ink-dot hands.
  • Bean head (h/w ~0.93, narrower chin, pointier crown).
  • No blush or eye highlights. Faces at brightness ~.93.
  • Walk bob 3-7 %, lean 1-7°, planted feet.
  • Blinks every 0.5-2.3 s, lasting 4 f.

Look-match loop (how the bridge look was reached; reuse for any "make it look like X" pass)

1. Measure the reference first. One analysis agent per topic (motion, depth, textures, characters, lighting, outlines, grade), each writing numbers, not adjectives, to references/reports/<topic>.md. Shared brief: compare/CONTEXT.md. Reports are the source of truth for every agent after.
2. Sandboxes, one owner per file. Copy the project folder to src/videos/<name>_{cam,world,mat,char,post} with entries src/sandboxes/<n>/index.tsx (same index.tsx as main, only the import path changed). Ownership used for the bridge:
   • cam: story.ts (camera keys, cuts, kicks, DOF table, foreground props)
   • world: world.ts, lights.ts, shade.ts, grade.ts (sun, sky, fog, river, colour, grade)
   • mat: paint.ts, build.ts (painted textures, buildings, props)
   • char: toon.ts, cast.ts, cardshade.ts (character drawing, posing, card lighting, fx cards)
   • post: ink.ts, Bridge.tsx, lensblur.ts (outline pass, composer, lens blur, bloom) Agents only edit their own files in their own sandbox and render stills with BRIDGE_ENTRY=src/sandboxes/<n>/index.tsx node scripts/bridge-stills.mjs out/<name>/<n>/x_ <frames> 0.5.
3. Merge with python scripts/bridge_merge.py [cam,world,mat,char,post] (backs main up to backup/bridge_<time>/, copies owned files only). Then npx tsc --noEmit (filter to the project folder), then re-sync every sandbox from main (cp -f src/videos/<name>/* src/videos/<name>_<n>/) before the next round.
4. Verify with a fresh Opus max agent: side-by-side sheets (python scripts/sbs_sheet.py out.jpg --ours a b --ref r1 r2), score each area out of 10 (colour day / golden, lighting, ink, textures, characters, depth, composition) plus a numbered fix list with frame numbers. Map each fix to the file owner and resume that owner's agent with SendMessage (keeps its context). Bridge scores: round 1 5/3/4/5/4/7/4/6 → round 2 5/4/5/6/4/7/6/6 → round 3 see notes.
5. Contracts between owners live in exported constants (e.g. NOINK layer from ink.ts, dofAt/FOREGROUND from story.ts, LIGHTS/LOOK from lights.ts). Changing a contract means telling every owner that reads it.

Gotchas

• Kenney nature-kit materials are teal: recolour them by material name in build.ts.
• Cameras end up inside houses: the sight-line culler in build.ts/world.ts removes anything blocking any camera ray.
• Ink in the same merged effect pass as DOF gets overwritten. Ink must be its own pass before the blur.
• Pixel-sized post parameters must scale with resolution: stills are at 0.5, the final at 1.333. R3F clamps dpr to ≥ 1, so --scale 0.5 stills still render a 1080x1920 buffer.
• Disk: every Remotion bundle copies public/ (~1.3 GB) into the temp folder\remotion-webpack-bundle-* and never deletes it. 60+ GB filled the system drive during the bridge passes and a render died with ENOSPC. Always render with export TEMP='<big drive>/remotion-tmp' TMP='<big drive>/remotion-tmp' (bridge-stills.mjs sets it itself) and ask the user before deleting old bundle folders in the shared the temp folder (auto mode blocks the sweep). /remotion-tmp fills too (~500 folders / 110 GB after the 4 bridge rounds): when a project is finished and nothing is rendering, delete its remotion-webpack-bundle-* and puppeteer_dev_chrome_profile-* folders there.
• In JS strings write "<big drive>/remotion-tmp", never "<big drive>\remotion-tmp" (\r becomes a carriage return).
• Remotion writes a silent audio track: mux with -map 0:v:0 -map 1:a:0.
• Sub-agents can't write report files in this harness: ask them for a ≤150-word final message and save it yourself to out/<name>/<n>/REPORT.md.
• Scratch .ts files inside out/ break tsc: name them .txt.
• 2K renders run ~4 fps (≈10 min for 67 s). Stills at 0.5 take 3-8 s each after a ~40 s bundle.
• Film grain at 2.5 % made the 2K file 910 MB instead of ~250 MB and will turn to mush on YouTube; the reference has none. Keep it off.

Readability traps the verifier caught (check every new prop/shot for these)

• A swing drawn as one rope that disappears behind the kid reads as a HANGING. Always two ropes to a wide plank with the kid between them (swingKid in toon.ts), and a frame whose legs clear the arc (A-frame at each end).
• Props or flotsam closer than ~2 m to a lens become huge cut-off blobs at the frame edge. When a camera moves, re-check every prop near its new position (the river cameras moved into the floating hats).
• A foreground bush cut by the frame edge reads as a "green ball"; a foreground prop must be recognisable or gone.
• Flags: red + yellow star/diamond reads as Vietnam's. Use a period-correct flag (St George's cross for England 1831).
• Painted signs lit by a side sun go grey-blue and unreadable: give the sign face an unlit material.
• Golden rim light done as an offset outer copy turns every hat into a picture frame: inner rim only, head/torso only.
• Balcony/railing rails at chest height through characters: put the lens ~2.5 m up so rails cross legs, not torsos.
• Flat holds read as freezes in QA (frozen span ≥12 f): every hold needs a creeping push-in.