⏳ This skill is pending AI review.
Scores will appear once the review pipeline completes.
business-motion-film
Plan, build and quality-check a premium short commercial for a real business (roofing, property, ecommerce, B2B software) using AI-generated footage, code-built motion (HTML/GSAP), selective Three.js and an independent-critic "Gauntlet" loop. Use when asked to make a launch-style / SaaS-style video, business ad, explainer, sample reel or pitch video, or to review and improve one. Encodes motion principles distilled from 28 professional launch films, a quality bar, audio rules and a business-offe
Choose how to use this skill
You do not need every option. Choose the path your AI client supports. The stable page stays the same; versioned files are immutable.
1. Native installer
This listing has no registered native installer command. Use the complete package or source fallback below, depending on what your client supports.
Do not guess an installer command or replace an existing version without reviewing the diff.
2. Complete package recommended
Download the ZIP when available. It includes SKILL.md plus the references, security notes and version metadata.
No complete ProSkills package is published for this listing yet.3. Prompt-only
Copy the prompt above when the agent can read the stable page or when you want to adopt the workflow without installing a skill.
Need only the instruction file?
Download SKILL.md only if your client requires a single file. The complete ZIP is safer for a full installation because it preserves the references and release context.
No path installs or executes anything by itself. Your agent still needs access to the project files. Before updating, compare the installed version and review the diff.
// RATINGS
Not yet listed on ClawHub or SkillsMP
// README
Motion Video Kit
A Claude Code skill (also usable as plain context for any LLM) for making premium, launch-style commercials for real businesses with AI-generated footage, code-built motion (HTML/GSAP), selective Three.js, and an independent-critic quality loop.
It packages what was learned from studying 28 professional SaaS launch films and from building three full sample commercials through dozens of rounds of independent critique: a calm service film, a 40 s Three.js product spec ad for a foldable phone in which every frame is code, and a 20 s smart-home display spec ad that a demanding client re-shaped for seven rounds after the critics said ship:
- The Gauntlet loop: builder ≠ judge, fresh critics on the actual render, item-by-item verification, and a ledger. Includes ready-to-use critic prompts.
- Motion grammar: six rules and a catalog of 16 reusable mechanisms, plus notes on all 28 reference films (links to the originals; no footage redistributed).
- Quality bar: measurable checks (frozen time, loudness, contrast, brand colour) and the visual and business criteria clients actually enforce.
- Audio rules: matching music to the buyer's customer, back-timing a real track so its own ending lands on the end card (one continuous piece, cuts on its beat), clean sound effects on every small move with the hero moments clearly audible, and mix targets.
- Business playbook: verticals with buying evidence and price anchors, a pilot-offer template, honesty rules.
- Three.js patterns: deterministic, seekable scenes; exploded layers with projected callouts; photos pinned in 3D context; animated option patches; a realism checklist.
- Product hero realism: how to match a real device's motion frame by frame (measured angle keys, a monotone cubic, locked camera), screen continuity and frost, deterministic accumulation motion blur, and lights that reveal the angle without flashing.
- Scripts and templates: frozen-time and loudness measurement, contact sheets, sound-effect softening, an isolated component lab, and a projected-overlay module.
Install (Claude Code)
git clone https://github.com/echris6/motion-video-kit.git
cp -r motion-video-kit/business-motion-film ~/.claude/skills/ # all projects
# or, per project:
cp -r motion-video-kit/business-motion-film <your-project>/.claude/skills/
Then ask for a business commercial, sample reel or explainer, or for a review of one, and the skill loads. It works best with HyperFrames for rendering, but the principles, prompts and checks are renderer-agnostic.
Other LLMs: paste business-motion-film/SKILL.md plus the reference files you need into the context.
Requirements for the scripts
ffmpeg/ffprobe (with the ebur128 filter). Optional: Node 22+ and HyperFrames for the lab template.
Layout
business-motion-film/
SKILL.md workflow + non-negotiables
references/
gauntlet.md the review loop, with real findings
critic-prompts.md component / full-film / verification / storyboard prompts
motion-grammar.md 6 rules + mechanism catalog + pacing numbers
launch-film-notes.md notes on 28 reference films (links only)
quality-bar.md measured + visual + business ship criteria
audio.md music, SFX, mix
business-offers.md verticals, evidence, price anchors, pilot offer
three-js-patterns.md render contract, patterns, realism checklist
case-study-alder.md a full worked example (calm service film)
product-hero-realism.md making a 3D device move like the real one
case-study-duo.md a Three.js product film, round by round
case-study-home-display.md what a client changed after the critics said SHIP
scripts/ frozen-time, loudness, contact-sheet, soften-sfx, sfx-candidates, solve-sfx-gains, offline-mix, backtime-music
templates/ component-lab.html, projected-overlays.js
Notes
- Reference films belong to their owners. This repo contains only original written analysis and links. Reach numbers quoted from studio portfolios are studio claims.
- Market prices are a September 2026 snapshot of public listings; re-verify before quoting.
- Never present AI-generated imagery as a real customer, job or result. The skill enforces labelling.
License
MIT for the text, scripts and templates in this repo.
// HOW IT'S BUILT
// FILES
24 files shipped with this skill. Inline text is the catalog copy. The upstream link is the source repository.
business-motion-film/SKILL.md7.9 KB05b7491127a1upstream ↗
--- name: business-motion-film description: Plan, build and quality-check a premium short commercial for a real business (roofing, property, ecommerce, B2B software) using AI-generated footage, code-built motion (HTML/GSAP), selective Three.js and an independent-critic "Gauntlet" loop. Use when asked to make a launch-style / SaaS-style video, business ad, explainer, sample reel or pitch video, or to review and improve one. Encodes motion principles distilled from 28 professional launch films, a quality bar, audio rules and a business-offer playbook. --- # Business motion film Make short (15–40s) commercials that look like premium SaaS launch films but sell a real business outcome: explain a service, earn a quote request, follow up on a quote, or refresh an ad. The visuals can be ambitious; the claims must be true. Read only the reference file your current step needs: | Step | Read | |---|---| | Choosing a vertical, offer, pilot scope, pricing anchors | `references/business-offers.md` | | Storyboarding and motion design | `references/motion-grammar.md`, then `references/launch-film-notes.md` for concrete moments | | Building 3D components | `references/three-js-patterns.md`, `templates/component-lab.html`, `templates/projected-overlays.js` | | A 3D product hero that must move like the real thing (unfolds, closes, reveals) | `references/product-hero-realism.md` | | Any review round | `references/gauntlet.md`, `references/critic-prompts.md`, `references/quality-bar.md` | | Music, sound effects, mix | `references/audio.md` | | A worked example end to end | `references/case-study-alder.md` (calm service film), `references/case-study-duo.md` (Three.js product film), `references/case-study-home-display.md` (what a client changed after the critics said SHIP) | ## Non-negotiables 1. **Truth.** Never invent testimonials, ratings, completed jobs, diagnoses, savings, warranties or confirmed appointments. A concept film says it is a concept ("Fictional brand · Concept film · AI-generated imagery"). A client film uses the client's approved footage and facts. 2. **Business purpose on mute.** A first-time viewer must be able to say what the business does and what to do next, with the sound off, by the end. 3. **Independent judging.** The builder never grades its own work. Every component and every full cut goes to a fresh critic that sees only the render, the brief and the references (`references/gauntlet.md`). 4. **Deterministic motion.** Every animated state is a pure function of timeline time: no free-running `requestAnimationFrame`, `Date.now()`, unseeded randomness or physics history. Seek to any frame and get the same pixels. 5. **"No bugs" is not "good".** Clean, smooth and error-free versions still get rejected for being basic, empty or slow. Measure (`scripts/`) and judge creatively. ## Workflow ### 1. Brief (15 min) Write `BRIEF.md`: business, buyer, the viewer's problem, the single action (CTA), what is provably true, available real assets, length, aspect ratios, brand (palette, type), and what must never be claimed. If the user has no real assets, agree it is a labelled concept. ### 2. References Pick 3–6 moments from `references/launch-film-notes.md` whose *mechanism* fits the story, e.g. a foreground fly-through, a persistent selected object, or a layered exploded view. If the user supplies reference videos, extract frames (`scripts/contact-sheet.sh`) and cite exact timestamps. Borrow mechanisms, never layouts, logos or footage. ### 3. Storyboard before building One table: time → what the viewer sees → business job of the beat → the transition out, including which object survives it. Rules: - 12–15 distinct compositions for a 30s film. Vary scale: macro → wide → overhead → UI close-up → type impact. - The lead subject fills 60–85% of the usable frame in feature beats. No small card floating in a big empty field. - Every beat has a readable landing, then an accelerating exit. No holds without motion over ~0.6s. - Name 3 signature transformations you can describe without effect names, e.g. "the inspection photo becomes the report's hero image". - 3D only where it explains something physical or spatial. Typography and UI stay crisp HTML/SVG. Get a quick critic pass on the storyboard itself (chronology, business logic, truthfulness). ### 4. Assets - **AI stills:** generate the key frame first, then animate it with image-to-video. Prompt as a production brief: lens, light, exposure ("bright, airy, nothing crushed to black"), negative space reserved for type, "no people / no text / no logos / no watermark" when needed. - **AI video:** short (4–5s) clips with restrained camera moves ("a steady gimbal dolly, about 25 cm of travel, preserve the exact geometry, no morphing"). Expect 720p/24fps and interpolate/upscale, which is not native detail. Reject clips whose geometry warps; regenerate rather than hide. - Keep a generation ledger: model, prompt, seed, job id, accepted window. Never commit API keys. ### 5. Build components in isolation Each custom component (3D scene, report UI, booking flow, logo mark) gets its own lab page (`templates/component-lab.html`), is rendered to stills plus a short motion proof, and goes through a component Gauntlet round **before** it joins the film. ### 6. Assemble One root timeline; each scene is a sub-composition; Three.js scenes are rendered from the root seek clock; overlays such as labels, pins and leader lines are positioned from projected 3D anchors every frame (`templates/projected-overlays.js`). Hand objects across scenes at *exact* pixel coordinates, so the last frame of scene A equals the first frame of scene B for the carried object. ### 7. Full-film Gauntlet Render the whole film. Run `scripts/frozen-time.sh` and `scripts/loudness.sh`. Send to a fresh critic with `references/critic-prompts.md` (full-film prompt). Fix the highest-impact items, re-render, and send to a *new* critic that verifies the previous list item by item. Stop at the quality bar or when gains become cosmetic (usually 3–5 rounds). ### 8. Audio last, but not an afterthought Follow `references/audio.md`. For picture-locked music, back-time a real track so its own ending lands on the end card and it plays as one continuous piece, with every cut on its beat (`scripts/backtime-music.py`); offer several tracks under the same picture and let the client pick by ear. Keep the music matched to the buyer's customer and well under the effects (lo-fi or warm acoustic for home services); one soft, rumble-free whoosh per real transition; clean UI sounds only on real actions; per-event levels solved in each effect's band (`scripts/solve-sfx-gains.py`); candidates screened with `scripts/sfx-candidates.py`. Always deliver a music-only fallback, and isolate complaints with it. When the client dislikes an effect, replace it with a cleaner one; don't delete it. Make the hero moments audible. ### 9. Deliver Master MP4 (1080p60 for motion work), poster frame, contact sheet, editable source, generation ledger, and a Gauntlet ledger (what each critic found → what changed → measured before/after). State limits honestly: what was measured versus listened to, and sampled versus exhaustive review. Give deliverables project-unique names, put each round in its own named folder, and open it for the client. Expect the client's own pass after the critics' SHIP, and turn it around by timestamp (`references/gauntlet.md`). ## Working style the clients in this playbook expected - Plan visibly (storyboard), then build without stopping to ask for approval of every step. - Report status in one line when working for a long time. - When time is cut, switch to deadline mode (`references/gauntlet.md`): hard stops, integrate, one critic, top fixes, deliver. - Never write API keys to files. If a key must be used, let the user run the one command in their own terminal, and suggest rotating any key pasted into a chat. - Show the work: contact sheets, measured numbers, critic verdicts.business-motion-film/scripts/backtime-music.py4.6 KB303a98c937b0upstream ↗
#!/usr/bin/env python3 """Back-time a music track to picture: its own final hit lands on the end-card cut, and it plays as ONE continuous piece. Nothing is cut, looped or spliced. The film opens mid-track; the track's real ending rings out under the end card. Optionally varispeed (tape-style: pitch moves with tempo) so its beat grid matches the film's cut grid. Usage: python3 backtime-music.py track.mp3 --list-hits # candidates for the final hit python3 backtime-music.py track.mp3 out.wav --film 18.5 --card 16.0 --hit 96.018 [--bpm 121 --to-bpm 120] [--fade-in 0.04] [--tail-ease 17.6:18.5] --hit the track's final-hit time in the source, in seconds (measure it: --list-hits, then confirm by ear or on a waveform) --bpm the track's measured tempo; with --to-bpm, varispeed by to/bpm (keep it within ~1.7 %, about 0.3 semitone) --tail-ease a:b ease the file end out from a to b (film seconds) if the ring-out is still sounding at the film's end Prints where the hit landed, the varispeed and the source window used. Needs ffmpeg, numpy and scipy. """ import argparse, subprocess, numpy as np from scipy.signal import resample_poly SR = 48000 ap = argparse.ArgumentParser() ap.add_argument('src'); ap.add_argument('out', nargs='?') ap.add_argument('--film', type=float); ap.add_argument('--card', type=float) ap.add_argument('--hit', type=float); ap.add_argument('--bpm', type=float); ap.add_argument('--to-bpm', type=float) ap.add_argument('--list-hits', action='store_true'); ap.add_argument('--fade-in', type=float, default=0.04); ap.add_argument('--tail-ease', default=None) a = ap.parse_args() raw = subprocess.run(['ffmpeg', '-v', 'error', '-i', a.src, '-ac', '2', '-ar', str(SR), '-f', 'f32le', '-'], capture_output=True, check=True).stdout x = np.frombuffer(raw, np.float32).reshape(-1, 2).astype(np.float64) def list_hits(x, search=20.0): """Strong attacks in the last `search` s, with how far the track drops in the 1.5 s after each. The button ending is usually a loud attack followed by a big drop, but a long ring-out can contain later, bigger-dropping attacks: auto-picking chose the wrong one on 3 of 6 library endings, so a person confirms by ear.""" m = x.mean(1); h = int(.01 * SR); n = len(m) // h r = 10 * np.log10((m[:n * h].reshape(n, h) ** 2).mean(1) + 1e-12); body = np.percentile(r, 90) rows = [] for i in range(max(3, n - int(search * 100)), n - 1): if r[i] > body - 9 and r[i] - r[i - 3:i].min() > 6 and r[i] >= r[max(0, i - 5):i + 6].max(): later = r[i + 5:i + 150]; drop = r[i] - (10 * np.log10((10 ** (later / 10)).mean()) if len(later) else -120) rows.append((i * h / SR, r[i] - body, drop)) print(' time s level re body drop over the next 1.5 s') for t, lv, dr in rows: print(f'{t:8.3f} {lv:+6.1f} dB {dr:5.1f} dB') print('Listen to the last few candidates before choosing: a ring-out can hold later attacks with bigger drops.') if a.list_hits: list_hits(x); raise SystemExit if a.hit is None: raise SystemExit('pass --hit (the source time of the final hit); --list-hits shows candidates') hit = a.hit ratio = (a.to_bpm / a.bpm) if (a.bpm and a.to_bpm) else 1.0 # >1 = faster if abs(ratio - 1) > 0.017: print(f'warning: varispeed {100 * (ratio - 1):+.2f} % is over 1.7 % and may be audible') if abs(ratio - 1) > 1e-6: up, down = 1000, int(round(1000 * ratio)); x = resample_poly(x, up, down, axis=0); hit = hit / ratio start = hit - a.card # source second playing at film 0 i0 = int(round(start * SR)); n = int(round(a.film * SR)) y = np.zeros((n, 2)); s0 = max(0, i0); d0 = s0 - i0; seg = x[s0:s0 + n - d0]; y[d0:d0 + len(seg)] = seg fi = max(1, int(a.fade_in * SR)); y[:fi] *= (0.5 - 0.5 * np.cos(np.linspace(0, np.pi, fi)))[:, None] if a.tail_ease: t0, t1 = (float(v) for v in a.tail_ease.split(':')); j0, j1 = int(t0 * SR), min(n, int(t1 * SR)) y[j0:j1] *= (np.cos(np.linspace(0, np.pi / 2, j1 - j0)) ** 2)[:, None]; y[j1:] = 0 y *= 10 ** (-1 / 20) / (np.abs(y).max() + 1e-12) # peak -1 dBFS; set the film level in the mix subprocess.run(['ffmpeg', '-v', 'error', '-y', '-f', 'f64le', '-ar', str(SR), '-ac', '2', '-i', '-', a.out], input=y.astype(np.float64).tobytes(), check=True) print(f'hit {hit * ratio:.3f} s in the source -> film {a.card:.3f} s; varispeed {100 * (ratio - 1):+.2f} % ' f'({12 * np.log2(ratio):+.2f} semitones); source window {start * ratio:.3f}-{(start + a.film) * ratio:.3f} s' + (' (the track starts after film 0: it is shorter than the film before its hit)' if i0 < 0 else ''))business-motion-film/scripts/contact-sheet.sh418 Bfdfc11aeec7eupstream ↗
#!/usr/bin/env bash # Contact sheet for critics: one frame every N seconds, tiled. # Usage: scripts/contact-sheet.sh film.mp4 out.jpg [every_seconds=1] [cols=6] [rows=5] [start=0] [duration=] set -euo pipefail f="$1"; o="$2"; e="${3:-1}"; c="${4:-6}"; r="${5:-5}"; ss="${6:-0}"; d="${7:-}" ffmpeg -loglevel error -y -ss "$ss" ${d:+-t "$d"} -i "$f" -vf "fps=1/$e,scale=480:-1,tile=${c}x${r}" -frames:v 1 "$o" echo "$o"business-motion-film/scripts/frozen-time.sh562 B84f87035f642upstream ↗
#!/usr/bin/env bash # Report near-frozen moments in a render: frame-to-frame luma difference below a threshold, sampled at 10fps. # Usage: scripts/frozen-time.sh film.mp4 [threshold=0.35] set -euo pipefail f="$1"; th="${2:-0.35}" ffmpeg -hide_banner -i "$f" -vf "fps=10,scale=320:-1,format=gray,tblend=all_mode=difference,signalstats,metadata=print:key=lavfi.signalstats.YAVG" -an -f null - 2>&1 \ | grep -o "YAVG=[0-9.]*" | awk -F= -v th="$th" ' {t=NR/10; if($2<th){n++; printf "%.1f ", t}} END{printf "\nnear-frozen samples: %d (≈%.1fs)\n", n, n/10}'business-motion-film/scripts/loudness.sh508 B93f65565a444upstream ↗
#!/usr/bin/env bash # Integrated loudness, loudness range, true peak, plus short-term loudness each second. # Usage: scripts/loudness.sh film.mp4 set -euo pipefail f="$1" out=$(ffmpeg -hide_banner -i "$f" -af ebur128=peak=true -f null - 2>&1) echo "$out" | grep -E "^ +(I|LRA|Peak):" | tr -s ' ' echo "short-term per second:" echo "$out" | grep -oE "t: *[0-9.]+ .*S: *[-0-9.]+" | awk '{for(i=1;i<=NF;i++){if($i=="t:")t=$(i+1); if($i=="S:")s=$(i+1)}; if(int(t*10)%10==0) printf "%d:%s ", t, s} END{print ""}'business-motion-film/scripts/offline-mix.py9.0 KB27ebb492bbe4upstream ↗
# offline-mix.py: mix a silent picture render with a music score and per-event sound effects, levelled in each effect's own band. # Project layout it expects (next to this script's parent, i.e. <project>/tools/offline-mix.py): # assets/music/<score>.mp3 assets/sfx/<name>.mp3 # assets/sfx/plan.json [[name, film_time_s, target_dB_over_music_in_band, note, (floor_dBFS), (peak_cap_dB)], ...] # Usage: python3 tools/offline-mix.py renders/picture.mp4 renders/out.mp4 --score my-score.mp3 [--no-sfx] [--music-lufs -29.9] # [--ring-db 0] [--hf-cap 4] [--mfloor -50] [--hfloor -50] [--pkcap 6] [--pkfloor -20] [--drop-dip -2.5] [--write-events] # The music chain (low shelf, two compressors, high shelf, loudness levelling), the post-chain dip before the drop (14.0-17.1 s: # edit for your film) and the ring-out lift after 37.2 s are film-specific defaults; change them to your cut points. # Needs: ffmpeg (ebur128), numpy, scipy. Writes review/mix-<out>.txt with per-event in-band lift, 150 ms body and 2-8 kHz lift. #!/usr/bin/env python3 """iPhone Duo audio: offline stem mix, measured per effect (the client's rule: subtle energetic bed ~-29 LUFS, each effect set in its own band a few dB over that quiet bed, never a 2-8 kHz spike). Usage: python3 tools/mix.py <picture.mp4> <out.mp4> [--score score-duo-a2-beat.mp3] [--no-sfx] [--music-lufs -29] Reads assets/sfx/plan.json: [[name, film_time, target_dB_over_bed_in_band, note], ...] Writes: out.mp4 (picture copied untouched + AAC 256k), review/mix-<stem>.txt (per-event report), assets/sfx/events.json (the solved linear gains, relative to the music's build gain, so build.py's HTML audio matches the delivered mix closely). Music chain (approved on AURELIN v4): -3 dB low shelf at 90 Hz, 3:1 compression, -3 dB high shelf at 5 kHz, then gain to the target integrated loudness; 1.2 s fade at the end. Effects: each placed at its time, its gain solved so its 50 ms in-band peak sits target dB over the music's in-band 50 ms peak in the same window; then 2-8 kHz lift capped at HF_CAP dB. """ import json, subprocess, sys, re, os import numpy as np from scipy.signal import butter, sosfilt from pathlib import Path P = Path(__file__).resolve().parents[1]; SR = 48000 args = sys.argv[1:]; pic, out = args[0], args[1] opt = lambda k, d: args[args.index(k) + 1] if k in args else d score = opt('--score', 'score-duo-a2-beat.mp3'); target_lufs = float(opt('--music-lufs', '-29.9')); HF_CAP = float(opt('--hf-cap', '4')) # the client's rule: 2-8 kHz lift <= ~4 dB no_sfx = '--no-sfx' in args def load(p): return np.frombuffer(subprocess.run(['ffmpeg', '-v', 'error', '-i', str(p), '-ac', '2', '-ar', str(SR), '-f', 'f32le', '-'], capture_output=True).stdout, np.float32).reshape(-1, 2).copy() def lufs(x): tmp = P / 'renders/_lufs.wav'; write(tmp, x) e = subprocess.run(['ffmpeg', '-hide_banner', '-i', str(tmp), '-af', 'ebur128=peak=true', '-f', 'null', '-'], capture_output=True, text=True).stderr return float(re.findall(r'I:\s+(-?[\d.]+) LUFS', e)[-1]), float((re.findall(r'Peak:\s+(-?[\d.]+) dBFS', e) or ['nan'])[-1]), float((re.findall(r'LRA:\s+([\d.]+) LU', e) or ['nan'])[-1]) def write(p, x): subprocess.run(['ffmpeg', '-v', 'error', '-y', '-f', 'f32le', '-ar', str(SR), '-ac', '2', '-i', '-', str(p)], input=np.clip(x, -1, 1).astype(np.float32).tobytes(), check=True) def pk(y): h = int(.05 * SR); return max(10 * np.log10((y[i:i + h] ** 2).mean() + 1e-12) for i in range(0, max(1, len(y) - h), h // 2)) N = int(40.0 * SR) # ---- music stem ---- raw = P / 'renders/_music_chain.wav' subprocess.run(['ffmpeg', '-v', 'error', '-y', '-i', str(P / 'assets/music' / score), '-af', 'lowshelf=f=90:g=-3,acompressor=threshold=-24dB:ratio=3:attack=4:release=140:knee=6:makeup=1,highshelf=f=5000:g=-3,acompressor=threshold=-30dB:ratio=2.5:attack=1:release=90:knee=4', # 2nd stage: music-critic-1 (A2 intro spikes 19/73 -> 0/73 windows >4.5 dB) '-ar', str(SR), '-ac', '2', '-t', '40', str(raw)], check=True) mus = load(raw)[:N]; mus = np.pad(mus, ((0, N - len(mus)), (0, 0))) L0 = lufs(mus)[0]; mg = 10 ** ((target_lufs - L0) / 20); mus *= mg fade = np.ones(N, np.float32); f0 = int((40 - 1.2) * SR); fade[f0:] = np.linspace(1, 0, N - f0) ** 1.5; mus *= fade[:, None] RING = float(opt('--ring-db', '7')) # music-critic-1: the ring-out after the 37.0 resolve sat 15-25 dB under the bed; lift it (100 ms ramp) r0 = int(37.2 * SR); ramp = np.ones(N, np.float32); ramp[r0:] = 10 ** (RING / 20); ramp[r0:r0 + int(.1 * SR)] = np.linspace(1, 10 ** (RING / 20), int(.1 * SR)); mus *= ramp[:, None] # the build into the 17.5 drop: a post-chain dip (the compressors flattened the pre-chain one; audio-critic-3: +0.2 LU → ~+2 LU into the drop) DIP = float(opt('--drop-dip', '-2.5')) if DIP: d0, d1, rr = int(14.0 * SR), int(17.1 * SR), int(.01 * SR); dg = np.ones(N, np.float32); lv = 10 ** (DIP / 20) dg[d0:d1] = lv; dg[d0:d0 + rr] = np.linspace(1, lv, rr); dg[d1 - rr:d1] = np.linspace(lv, 1, rr); mus *= dg[:, None] mono_m = mus.mean(1) # ---- effects ---- fx = np.zeros_like(mus); rep = []; ev_out = [] plan = [] if no_sfx else json.loads((P / 'assets/sfx/plan.json').read_text()) HB = butter(4, [2000, 8000], btype='band', fs=SR, output='sos') PKCAP = float(opt('--pkcap', '6')); PKFLOOR = float(opt('--pkfloor', '-20')); MFLOOR = float(opt('--mfloor', '-50')); HFLOOR = float(opt('--hfloor', '-50')) build_gain = float(os.environ.get('DUO_SCORE_GAIN', '0.35')) for ev in plan: name, t, target, note = ev[:4]; floor = ev[4] if len(ev) > 4 else MFLOOR # optional per-event floor (dBFS in-band peak) pkc = ev[5] if len(ev) > 5 else PKCAP # optional per-event peak cap (dB over the local music peak) s = load(P / 'assets/sfx' / f'{name}.mp3'); m1 = s.mean(1) F = np.abs(np.fft.rfft(m1 * np.hanning(len(m1)))) ** 2; f = np.fft.rfftfreq(len(m1), 1 / SR); c = np.cumsum(F) / F.sum() lo = max(f[np.searchsorted(c, .2)], 60); hi = min(max(f[np.searchsorted(c, .8)], lo * 2), SR / 2 - 500); sos = butter(4, [lo, hi], btype='band', fs=SR, output='sos') i = int(t * SR); W = min(len(m1), int(.4 * SR)); seg_m = mono_m[max(0, i - SR // 2):i + W + SR // 2] mpk = pk(sosfilt(sos, seg_m)[SR // 2:SR // 2 + W]); epk = pk(sosfilt(sos, m1)[:W]) # floors: over near-silent music (a dead stop, a sparse intro) "+N dB over the music" would make an effect inaudible, so the # reference never drops below MFLOOR in-band, and the 2-8 kHz cap is measured against at least HFLOOR (Duo, lib cues) mpk_ref = max(mpk, floor) if os.environ.get("MIXDBG"): print(f"DBG {t:6.2f} mpk {mpk:.1f} hf {pk(sosfilt(HB, seg_m)[SR // 2:SR // 2 + W]):.1f}") g = 10 ** ((mpk_ref + target - epk) / 20) # 2-8 kHz cap: the lift of (music + effect) over music in the ear-sensitive band def hf_lift(g): a = seg_m.copy(); k = min(W, len(seg_m) - SR // 2); a[SR // 2:SR // 2 + k] += g * m1[:k] return pk(sosfilt(HB, a)[SR // 2:SR // 2 + W]) - max(pk(sosfilt(HB, seg_m)[SR // 2:SR // 2 + W]), HFLOOR) hl = hf_lift(g) while hl > HF_CAP and g > 1e-4: g *= .85; hl = hf_lift(g) # peak cap (Duo, client: "too loud"): the effect's sample peak may sit at most PKCAP dB over the music's local peak (floored) k0 = min(len(m1), len(seg_m) - SR // 2); mloc = max(20 * np.log10(np.abs(seg_m[SR // 2:SR // 2 + W]).max() + 1e-9), PKFLOOR) while 20 * np.log10(g * np.abs(m1[:k0]).max() + 1e-9) > mloc + pkc and g > 1e-4: g *= .9 j = min(N, i + len(s)); fx[i:j] += g * s[:j - i] e = np.zeros_like(seg_m); k = min(len(m1), len(seg_m) - SR // 2); e[SR // 2:SR // 2 + k] = g * m1[:k] inb = pk(sosfilt(sos, seg_m + e)[SR // 2:SR // 2 + W]) - mpk # film-critic-3: also report BODY, the 150 ms in-band RMS of (music + effect) over music alone (what a listener hears as weight) b0, b1 = SR // 2, SR // 2 + int(.15 * SR); rms = lambda y: 10 * np.log10((y[b0:b1] ** 2).mean() + 1e-12) body = rms(sosfilt(sos, seg_m + e)) - rms(sosfilt(sos, seg_m)) rep.append(f'{t:6.2f} {name:14s} band {int(lo):5d}-{int(hi):5d} Hz target +{target:.0f} in-band lift {inb:+5.1f} dB body {body:+5.1f} dB 2-8k lift {hl:+5.1f} dB gain {g:.4f} {note}') ev_out.append([name, t, round(len(s) / SR + .05, 3), round(float(g) / mg * build_gain, 4)]) mix = mus + fx peak = np.abs(mix).max(); lim = 10 ** (-1.5 / 20) if peak > lim: mix *= lim / peak # safety only; the mix is quiet by design wav = P / f'renders/_mix-{Path(out).stem}.wav'; write(wav, mix) subprocess.run(['ffmpeg', '-v', 'error', '-y', '-i', pic, '-i', str(wav), '-map', '0:v:0', '-map', '1:a:0', '-c:v', 'copy', '-c:a', 'aac', '-b:a', '256k', '-movflags', '+faststart', out], check=True) Lm, _, _ = lufs(mus); Lx, tp, lra = lufs(mix) hdr = f'score {score}: music-only {Lm:.1f} LUFS (chain gain {20*np.log10(mg):+.1f} dB); mix {Lx:.1f} LUFS, LRA {lra}, true peak {tp} dBFS; {len(plan)} effects' (P / 'review').mkdir(exist_ok=True); (P / f'review/mix-{Path(out).stem}.txt').write_text(hdr + '\n' + '\n'.join(rep) + '\n') if plan and '--write-events' in args: (P / 'assets/sfx/events.json').write_text(json.dumps(ev_out, indent=0)) print(hdr); print('\n'.join(rep))business-motion-film/scripts/sfx-candidates.py1.2 KBbded81d83cc7upstream ↗
#!/usr/bin/env python3 """Rank sound-effect candidates before anyone has to listen. Usage: scripts/sfx-candidates.py file1.mp3 file2.wav ... Reports active duration, rise time, spectral centroid, share of energy below 150 Hz (rumble/boom) and above 6 kHz (hiss/click), and a verdict. Reject: rumble > 50%, highs > 40% (for UI sounds), very long whooshes (> 0.8s) for repeated transitions.""" import subprocess, sys, numpy as np SR = 44100 for f in sys.argv[1:]: y = np.frombuffer(subprocess.run(['ffmpeg','-v','error','-i',f,'-ac','1','-ar',str(SR),'-f','f32le','-'],capture_output=True).stdout, np.float32) if not len(y): print(f, 'unreadable'); continue env = np.sqrt(np.convolve(y**2, np.ones(441)/441, 'same')); pk = env.max() act = np.where(env > pk*.05)[0]; dur = (act[-1]-act[0])/SR; rise = (np.argmax(env)-act[0])/SR*1000 F = np.abs(np.fft.rfft(y))**2; fr = np.fft.rfftfreq(len(y), 1/SR); tot = F.sum() cen = (F*fr).sum()/tot; lo = F[fr < 150].sum()/tot; hi = F[fr > 6000].sum()/tot why = [w for w, bad in [('boomy', lo > .5), ('hissy/clicky', hi > .4), ('long', dur > .8)] if bad] print(f"{f}: {dur:.2f}s rise {rise:.0f}ms centroid {cen:.0f}Hz <150Hz {lo*100:.0f}% >6k {hi*100:.0f}% -> {'REJECT: '+', '.join(why) if why else 'ok'}")business-motion-film/scripts/soften-sfx.sh565 B16bdcc1d162eupstream ↗
#!/usr/bin/env bash # Soften a library sound effect: high-pass, low-pass, click-free fades. # Usage: scripts/soften-sfx.sh in.mp3 out.mp3 [highpass=150] [lowpass=6500] [fade_out_start] [fade_out_dur=0.15] set -euo pipefail in="$1"; out="$2"; hp="${3:-150}"; lp="${4:-6500}" dur=$(ffprobe -v error -show_entries format=duration -of csv=p=0 "$in") fd="${6:-0.15}"; fs="${5:-$(awk -v d="$dur" -v f="$fd" 'BEGIN{print d-f}')}" ffmpeg -loglevel error -y -i "$in" -af "highpass=f=$hp,lowpass=f=$lp,afade=t=in:d=0.01,afade=t=out:st=$fs:d=$fd" -ar 48000 "$out" echo "$out"business-motion-film/scripts/solve-sfx-gains.py2.0 KB3ea643742b99upstream ↗
#!/usr/bin/env python3 """Solve a per-event gain so each effect lifts TARGET dB inside its own frequency band over the music, while the ear-sensitive 2-8 kHz band never lifts more than CAP dB. Usage: scripts/solve-sfx-gains.py music_only.mp4 events.json [target=3.5] [cap=4] events.json: [["sfx/whoosh.mp3", 2.67], ["sfx/pop.mp3", 10.9], ...] (file, start time in seconds) Prints [file, time, gain] rows. Override collapsed gains for repeated sounds so they stay consistent.""" import subprocess, sys, json, numpy as np from scipy.signal import butter, sosfilt SR = 44100 def load(p): return np.frombuffer(subprocess.run(['ffmpeg','-v','error','-i',p,'-ac','1','-ar',str(SR),'-f','f32le','-'],capture_output=True).stdout, np.float32).copy() def peak(y): h = int(.05*SR); return max(10*np.log10((y[i:i+h]**2).mean()+1e-12) for i in range(0, len(y)-h, h//2)) music = np.concatenate([np.zeros(SR, np.float32), load(sys.argv[1]), np.zeros(SR, np.float32)]) events = json.load(open(sys.argv[2])); target = float(sys.argv[3]) if len(sys.argv) > 3 else 3.5; cap = float(sys.argv[4]) if len(sys.argv) > 4 else 4.0 HB = butter(4, [2000, 8000], btype='band', fs=SR, output='sos'); W = int(.3*SR) for path, t in events: s = load(path); F = np.abs(np.fft.rfft(s*np.hanning(len(s))))**2; f = np.fft.rfftfreq(len(s), 1/SR); c = np.cumsum(F)/F.sum() lo = max(f[np.searchsorted(c, .2)], 40); hi = min(max(f[np.searchsorted(c, .8)], lo*2), SR/2-200) sos = butter(4, [lo, hi], btype='band', fs=SR, output='sos') i = int(t*SR)+SR; base = music[i-SR//2:i+SR+SR//2] b0 = peak(sosfilt(sos, base)[SR//2:SR//2+W]); h0 = peak(sosfilt(HB, base)[SR//2:SR//2+W]); g = prev = .005 for x in np.geomspace(.005, 1.2, 140): seg = base.copy(); seg[SR//2:SR//2+len(s)] += x*s[:len(seg)-SR//2] if peak(sosfilt(HB, seg)[SR//2:SR//2+W]) - h0 > cap: g = prev; break if peak(sosfilt(sos, seg)[SR//2:SR//2+W]) - b0 >= target: g = x; break prev = x print(json.dumps([path, t, round(float(g), 3)]))business-motion-film/references/audio.md12 KB9fd1db1f4c01upstream ↗
# Audio Audio made otherwise strong visuals feel amateur more than anything else. These rules came from direct client rejections. ## Music - **Match the music to the buyer's customer, not to "tech launch".** An energetic SaaS-launch score was rejected for a roofing/homeowner film. For home services, warm, chill, acoustic-leaning music fits: felt piano, clean or nylon guitar, a soft steady pulse (brushed kit, shaker), ~84–88 BPM, major key. - **Chill still needs a pulse.** Beatless ambient or very sparse tracks feel sleepy against fast cuts and need heavy gain to be heard. - **Structure the track to the film.** With generative music (e.g. ElevenLabs composition plans), write one section per scene group with exact durations that sum to the film length. Minimum section length is often 3s, so merge short sections. Ask for "starts immediately on the first beat, no long intro" and "clear resolved final chord at the start of the last section, then natural ring-out". - **Check the generated track's loudness over time** (short-term per second). Reject tracks with a near-silent intro, dips under key scenes, or decay seconds before the logo. Regenerate with a different seed; some styles (e.g. tremolo electric guitar + Rhodes) repeatedly produced slow intros. - **Avoid:** vocals/humming, lo-fi vinyl crackle, ukulele/whistling corporate stock, trailer hits, EDM drops, risers, anything "epic". - **Deliver 2–3 options** cut to the same picture at identical integrated loudness, so the client compares music, not volume. ## Sound effects What the reference launch films do (audio analysis of 14 motion-first films): the music carries the sound design, with punchy rhythmic beds and cuts on beats. Effects are whooshes, pitched sweeps, sub hits on hard cuts, and crisp UI/product clicks. That density suits punchy tech tracks. **Copying it onto a calm score was rejected** as "too loud… not clean… annoying throughout". What a client finally approved for a calm home-services film, after eight rounds: - **One short, soft, rumble-free whoosh on every real transition:** scene changes, visible wipes, and the headline slam. Never long, boomy "air/wind" whooshes. Choose whooshes with little energy below 150 Hz and a gentle ~80–400 ms rise. The rejected ones had 35–83% of their energy below 150 Hz and sounded like a repeated boom. - **Small clean UI sounds only on real actions:** clicks on cursor clicks, pops on pins/badges, light ticks on labels, a check when something seats, one soft chime on confirmation, tiny taps when a logo locks together. - **3D reveals get a musical accent:** a clean rising three-note pluck as roof layers lift apart, and soft wooden taps as new tiles land. Pitch-shift tonal accents to the score's key. - Music chosen: lo-fi (Rhodes + soft dusty drums), about 30% quieter than first mixed, so effects read clearly. Rules: 1. **Clean sources.** A standard library (Mixkit Sound Effects Free License, HyperFrames' bundled SFX) with a gentle high-pass (~150–220 Hz), low-pass (~9–11 kHz) and click-free fades. Loosely prompted AI effects and synthesized sine sweeps/ticks were rejected. Targeted AI effects work when you **generate several candidates and reject by analysis** (`scripts/sfx-candidates.py`): drop anything boomy (>50% below 150 Hz), hissy or clicky (>40% above 6 kHz), or noise-like (it will read as a whoosh). 2. **Set levels per event, in each effect's own frequency band,** against a music-only render (`scripts/solve-sfx-gains.py`). Target about +3–4 dB in-band, and cap the ear-sensitive 2–8 kHz lift at ~4 dB. Full-band loudness is useless here: effects barely move it, yet ticks can spike 12–17 dB at 2–8 kHz and sound annoying. 3. **Keep repeated sounds consistent.** When the solver collapses a gain because the music is momentarily quiet, override it to match the sound's other occurrences. Soften sounds that cluster within ~0.15s (×0.6) so they don't stack. 4. **Isolate complaints with a music-only render.** When the client says "that whoosh", first check whether it's in the music (brushed snares and risers sound like whooshes). Ask them to play the music-only version at that moment, then fix the right layer. 5. **Change one thing per round and name it by timestamp.** The window between "I can't hear anything" and "I want to break my laptop" is a few dB. 6. **Always deliver a music-only fallback,** and keep previous versions in an `older versions/` folder. 7. If the music resolves on the logo, don't add a separate musical sting (key clashes), and don't cut the music out before the logo on calm pieces. ## Mix & master - Mix and render as part of the composition (per-track volume, automation for dips under dense reading moments). - **Loudness:** reference launch films master around −14 LUFS (some at −7 to −10). The approved calm film landed at −19 LUFS integrated once the client lowered the music 30%, with effects clearly audible on top. A dense effects layer at −14 was rejected as too loud. True peak ≤ −1 dBFS with a limiter. Copy the video stream untouched; AAC 256k. - Measure with `scripts/loudness.sh`. State honestly whether anyone actually listened, since measurements can't judge taste. ## Energetic product films (from a 40 s Three.js product spec, 2026-09) - **Library music beat the AI scores.** The client called two rounds of generated "brand-style" scores "a bit off". A human-made library track, **edited to picture**, landed well. Mixkit stock music is reachable at `https://assets.mixkit.co/music/<id>/<id>.mp3` (IDs come from the `data-audio-player-preview-url-value` attributes on the tag pages). Check the licence and credit the tracks. - **Edit it like a music editor.** The next client rejected the dead stop and the spliced hit as "unnecessary sudden cuts", so prefer back-timing (next section) and use these edits only when a client wants a drop: - time-stretch to a tempo where every cut falls on a beat (120 BPM puts cuts at x.0 and x.5); - measure the real onset at each cut and stretch ≤0.5 % between cuts (a generic beat-warp locked onto the off-beat); - a sparse or filtered intro, with the groove arriving on the hero moment; - a dead stop (about 0.4 s) and the track's own drop on the biggest cut; - the track's own final hit spliced on the end card, with equal-power crossfades on bar lines; - a 5 ms fade-in, because a full-level first sample clicks. - **Make the drop land.** A drop that measured only +0.2 LU needed two things: a 40→300 Hz high-pass build over the 2 s before it, and a −2.5 dB dip applied **after** the bus compressors (before them, the compressors flatten it out). Result: +2.7 LU, and the bass +9 dB into the drop. - **Brand study, for grammar only.** The reference brand film measured about 124 BPM, four-on-the-floor, with flat and confident density, a restrained 2–5 kHz, dead stops and a button ending. Don't name brands or artists in generator prompts; they get rejected. - **Levels the client approved:** - music at about −30 LUFS (he asked for 10 % less than −29); - whooshes as a soft "air pass" at about +2–3 dB in-band, raising short-term loudness ≤0.5 LU. A bundled whoosh at +7 dB was "too loud" and unprofessional; - micro sounds on **every small mechanical move** (screws backing out, parts re-seating, detents, panes docking): real recordings, each peaking **at or below** the local music peak. - **Mixer guards** (`scripts/offline-mix.py`): - **Level floors:** over a dead stop or a sparse intro, "+N dB over the music" makes an effect inaudible. A per-event floor on the reference level fixes it. - **A high-frequency floor:** the 2–8 kHz cap is measured against a floor, or a bass-only intro lets nothing through. - **A peak cap:** no effect may peak more than about 6 dB over the local music peak (8 dB for a signature hit). Without it, a key-press thock spiked to −2.8 dBFS in a −29 LUFS mix. - **Report "body" as well:** the 150 ms in-band RMS lift, verified on the decoded file against a music-only render. The 50 ms peak alone over-reported a 0.15 s thock by 8 dB. - **Synthesised thocks were rejected** as "a pitched tom" or not professional. Layered real recordings worked, e.g. a laptop key press plus a soft body punch, about 100 ms long. ## Picture-locked music without edits: back-timing (from a 20 s smart-home spec ad, 2026-10) The music editor's cut from the section above (a dead stop before the drop, the final hit spliced onto the end card) was rejected on the next film as **"unnecessary sudden cuts"** and "doesn't match". So were two rounds of generated scores timed to the picture ("still not the brand's style"). What the client approved: 1. **Real tracks with a real ending.** Screen a library for tracks near the film's tempo whose ending is a clean button or final hit with a short ring-out. Measure the ending; don't trust the title. 2. **Back-time each one.** Line the track up so its own final hit lands on the end-card cut (to the frame), then play it **as one continuous piece** from film 0. Nothing is cut, looped or spliced. The film opens mid-track, which is how the brand's own films open. `scripts/backtime-music.py` does it. 3. **Make the beat match by moving the picture, not the music.** Put every cut on the track's beat grid. At 120 BPM, cuts land on x.0 and x.5 s. Move the end-card cut onto the grid instead of chopping the track to fit an off-grid cut. If the track's tempo is off, varispeed it to the grid (tape-style; ≤ 1.7 %, ≈ 0.3 semitone, is inaudible). If the final hit is an off-beat "push", put the hit on the cut anyway: the other cuts then land on off-beats, which are still on the 8th-note grid. 4. **Let the client pick by ear.** We can't listen. Offer 5–7 back-timed tracks mixed under the **same** picture, at identical loudness and with the effects re-levelled per track (see below). Name the files by track. The client picked from that set in one pass after five rounds of single proposals. 5. **Anything driven by the music** (a woofer that pumps to the bass, lights on the beat) must be re-driven from the chosen track. It also needs to know its own start time in the film, so that re-timing the film doesn't break it. **Loudness.** −30 LUFS was "too loud" for this film, where a previous energetic film was approved at −29/−30. −34 LUFS was approved. Start a new film around −32 and let the music-only render settle it. **Generated music (ElevenLabs Music).** Composition plans timed to the picture worked mechanically: one take landed its final hit within 150 ms of the end card on the first try. But the client judged four style rounds "not the brand's style". Tips: - don't name brands or artists in prompts, or the request is refused; - describe the measured grammar instead (tempo, kick pattern, claps on 2 and 4, density, a button ending); - respect the plan's concurrency limit (2 parallel requests here; retry 429s after 30–40 s). ## Effects: what a demanding client asked for after "every small move needs a sound" (2026-10) - **Clean, one family.** For numbers counting and screen state changes, use a single soft UI click family, very subtle: about 0–1 dB in-band, "barely there". A textured "real" sound (a wood rattle for blinds closing) and a swoosh on a reveal were called "random". **Replace a disliked sound with a cleaner one; don't remove it.** The client noticed the removals and asked for them back. - **Hero mechanical moves must be heard.** A display tilting or a token landing, the moments the film is about, ended up at **+8–10 dB in-band**. The "peak at or below the local music peak" micro rule had capped them 2–4 dB under the music's peak: "maybe we can't even hear it". Give these events an explicit cap and exempt them from the micro rule. - **A feature gets its own signature sound.** For "room-filling spatial audio": a clean sub knock (HP 35 Hz / LP 500 Hz) as the speaker opens, plus a short swell widened in stereo (side ×1.6) that peaks as the title lands. Both are real recordings. The client asked for it, then for "a little less": it settled at about +4.7 dB in-band. - **Re-level the effects for each music option.** Caps computed against one track are wrong for another. When an effect lands in a staccato track's micro-gap, its "in-band lift" can read +15–20 dB while its absolute gain is tiny. Check the gain and the 2–8 kHz lift before calling it loud.business-motion-film/references/business-offers.md4.2 KBb9f76deb9037upstream ↗
# Selling motion films to businesses Research snapshot: September 2026, US market. Prices are public listings and case studies. They show that buying categories exist, not that any buyer will purchase *your* offer. Re-verify before quoting. ## Positioning - Sell the **business job a video does**: explain the service, earn quote requests, follow up on quotes, refresh ads monthly. Don't sell "AI video". - Target buyers who **already spend on marketing and already have footage**: established contractors, multi-property operators, and the agencies serving them. - The client's real footage is the proof. AI-generated imagery is for concepts, supporting shots and pitches, always labelled. - The commodity tier is crowded ($30 AI tools, $150 reels). Differentiate on judgment, accuracy to the client's real business, quality control, and fast variants. ## Verticals | Vertical | What to sell | Buying evidence | Caveats | |---|---|---|---| | **Roofing / home services** | Project stories, process explainers, objection handling, quote follow-up, monthly ad variants | Roof Lens: 10 short videos for $1,500/month from client footage, incl. publishing, quarterly commitment (rooflens.com/pricing). Tasca Studios: $7,500 roofing project incl. filming, drone and graphics (tascastudios.com/video-case-studies/neumann-roofing) | The package price isn't transaction volume; the $7.5k project includes filming | | **Apartment operators / property agencies** | Floor-plan explainers, leasing ads, amenity videos, availability/offer variants | Legend Management commissioned 285 unit tour videos (Grace Hill/Realync case). Conversion Logix lease-up campaign used YouTube + social video (48 leases, multi-channel). Focus Media lists a +$150 property-reel add-on | Tours were filmed; the lease result isn't video-only. Sell to multi-property operators or agencies, not single agents | | **Ecommerce** | Product demos, feature animations, product-page clips, ad-hook variants | Soona: 15–30s product highlight video $455/SKU + studio fee. Creatify/Vitals: $30 AI add-on, 1,000+ creatives/month (vendor-reported) | Cheap AI substitutes exist; generic image animation won't stand out | | **B2B software** | Sales explainers, onboarding, feature updates, customer education | Vidico/Brightflow (explainer + onboarding + brand). HoneyBook reported +9% waitlist signups in an A/B test (vendor-hosted) | Needs real product access; best fit for fully code-built motion | | **Solar** (later) | Installation-process or FAQ explainer | Boston Solar/Energy Circle: one 90s film cut into three short ads (2019–20) | Historical; savings claims need current, local, approved numbers | A good first test is roofing/home services: a close competitor offer exists to price against, the buyer is clear, the sales use is concrete (help a homeowner understand the job before a quote), and there are no financial claims. For a showcase built entirely in code, a B2B software workflow explainer is the easiest fit. ## Pilot offer template - One core 20–30s film + two opening hooks + two aspect ratios (16:9, 9:16) = four exports. - One consolidated revision round. - The client supplies approved footage, service area, process facts and CTA. Filming, ad spend, publishing and campaign management are separate line items. - Time your real production and revisions before setting a price. - First success test: does the buyer pay and use it? Lead quality and booked jobs need their campaign/CRM data later. ## Price anchors (Sept 2026) $30 AI add-on · $150 property reel · $455 product clip · $1,500/month for 10-video retainer · $7,500 custom project with filming. The defensible middle is probably recurring, business-specific creative, not one-off generic clips (a hypothesis to test). ## Honesty rules for sales material - Never present AI imagery as a real completed job, customer or result. - No invented testimonials, ratings, savings, warranties or performance lifts. - Vendor case-study numbers are vendor claims; say so if you cite them. - Label concept films as concept films. ## Unknowns to validate with real buyers Customer acquisition cost, close rate, production margin with client footage, whether a remote motion offer beats local videographers or in-house phone footage, and which vertical converts best.business-motion-film/references/case-study-alder.md4.6 KB0f0c3f4d8c04upstream ↗
# Case study: ALDER (fictional roofing inspection brand) A worked example of the whole playbook: one working day, two coding agents (Codex, then Claude Code), HyperFrames + GSAP + Three.js, Replicate (gpt-image-2.5-flare stills, Seedance 2.5 video), ElevenLabs (music), and a bundled SFX library. ## Brief Sell a roof inspection: a careful inspection turns small details into clear next steps. CTA: "Request a roof inspection". Tagline: *Small details. Big decisions.* Palette: forest `#0a211d`, cream `#f4f0e5`, citron `#d9ef71`; Manrope. Labelled "Fictional brand · Concept film · AI-generated imagery". No diagnosis, savings, guarantee or confirmed appointment is claimed; the form says the company "will confirm availability". ## Final structure (28.6s, 1080p60) | Time | Beat | Technique | |---|---|---| | 0–2.75 | SMALL DETAILS → BIG DECISIONS | Bright sunlit macro (AI still → image-to-video → 1080p60), forest type on sky with a citron bar, citron band wipe | | 2.75–6.05 | Built in layers | Three.js maquette house; front roof plane separates into shingles / underlayment / deck / rafters; four evenly spaced projected callouts | | 6.05–10.4 | Surface · Junctions · Drainage | Three AI inspection clips, one title position, one right-to-left wipe with a citron edge, focus reticles, contact strip | | 10.4–14.5 | Every photo. In context. | Camera flies around the same house; each photo pops out of the spot it documents, docks into a rail, then lands at the report's exact coordinates | | 14.5–17.6 | Your roof, explained. | Report with equal 124px rows; thumbnail click swaps the photo; an anchored flashing callout resolves into the Materials row; slow push | | 17.6–20.8 | Talk through options. | The discussion row heading flies to become the scene heading; the selected photo parks in the corner; 3D maintain/repair/replace patches | | 20.8–23.1 | Start with a visit. | Compact form; a visible Morning choice survives into the "Visit requested" confirmation | | 23.1–24.9 | Confidence overhead. | Daylight home, slow push | | 24.9–28.6 | ALDER | Bevelled 3D mark assembles, moves into the lockup, letters rise flat, CTA | ## How it evolved (and why) 1. **10s restrained ad** (image → video → text) was rejected as "too basic". The client wanted SaaS-launch density. 2. **30s and 40s cuts** added a storyboard, custom UI and a Three.js roof slab. They were rejected: the generic roof mesh and an evening house were overused, and it "looks like a basic video". 3. **36s cut** fixed spacing and card-size complaints but kept small 3D objects in large empty fields. The critic measured ~10s of frozen time. 4. **v8 (31s → 28.6s)** replaced the slab with the maquette house (layers + photo-pin flight), added 3D option patches and a new bright opener, and cut every hold. Five critic rounds brought frozen time down to ~0.7s. 5. **Audio (eight rounds):** energetic score rejected ("doesn't match; chill and mellow") → AI effects rejected ("not professional, jarring") → sparse library effects inaudible ("I can't hear anything") → launch-style dense layer rejected ("too loud, annoying") → rumbly repeated whooshes rejected ("whoosh, whoosh, whoosh") → one whoosh identified as the culprit by a music-only A/B → roof split too quiet → final: Lofi score 30% quieter, one soft rumble-free whoosh per transition, clean UI clicks/pops/taps, a key-tuned pluck on the 3D roof split, per-event in-band levels, −19 LUFS. ## Component critic findings that mattered - Layers: one-frame pop, a camera ending too tight, label collisions, toy gutters, flat glass, grey sky instead of brand cream. - Photo map: a pin on the wrong face of the chimney, a floating flashing "plinth", a grey gutter model versus a white gutter in the photo, a top-down slab moment. - Options: an unclipped sweep, two tiles airborne at once, end states that looked alike, parts leaving their cards. ## Generation volume 10 AI stills and 10 AI video clips (5 used in the final film), 7+ music scores, and 14 generated sound effects that were eventually all replaced by library effects. Each full 1080p60 render took ~1 minute locally. Track API spend from the start; it wasn't tracked here. ## Lessons - The strongest upgrade was making 3D *explain* (how the roof is built, where each photo was taken), not adding more 3D. - Carrying objects across scenes at exact pixel coordinates made the film feel "one continuous thing". - Independent critics found defects the builder repeatedly missed. Measurement (frozen time, loudness, sampled colours) turned taste arguments into fixable numbers. - Audio choices must follow the viewer (homeowners), not the genre of reference (tech launches).business-motion-film/references/case-study-duo.md3.7 KBaa207c44f030upstream ↗
# Case study: a 40 s foldable-phone spec commercial (Three.js-led) An unofficial concept ad for a foldable phone, made in the brand's visual language. It is local only, and the end card carries an "unofficial concept, not affiliated" disclaimer. Every frame is code: six Three.js canvases on one HyperFrames seek clock, with HTML/SVG type on top. The music and effects are licensed library audio, and the wallpapers and photos are AI-generated stills. ## Structure - **Three acts:** - dark, 0–17.5: the unfold, the hinge x-ray, the materials macro and the thinness shot; - light, 17.5–36: screen sizes, Split View, the camera, the x-ray, the colours; - dark, 36–40: a magnetic close into the end card. - **Handoffs:** - 7.0 and 14.0: hard cuts on the downbeat, pixel-registered; - 17.5: the edge's glint line opens into a band of light that *is* the light stage; - 25.0: a mid-turn cut on a shared function; - 32.5: a pixel-matched pose; - 35.75: the next scene arrives over the previous one's close-up. - One shared device module. A fold kinematic `pose({open, bias})` drives both halves, a spine knuckle and the displays, whose textures are built with canvas 2D. ## Process - **Gauntlet:** every component went through lab → proof render → fresh critic → fix → re-verify → integrate on KEEP. That covered the device, UI, screens, hinge, optics, colours, finale, the type system, images, effects and music. - **Rounds:** about 70 critic rounds in total, including 6 full-film rounds and 2 fold-specific critics comparing against the reference film. - **The unfold:** it took 5 builder rounds after the frame-by-frame study (`product-hero-realism.md`). - **The measured bar at ship:** - frozen time 0.2 s; - frame 0 a finished composition; - no text collisions and type inside title-safe; - no near-black frames on the light stage; - facts exact, and `hyperframes check` with 0 errors; - mix at −30 LUFS with a true peak of −12.9 dBFS. ## What the critics caught that the builders missed - A mid-swing speed hiccup caused by blending measured keys with a smoothstep. - The outer display sleeping too early: the director's reading of the reference was wrong. - Three pure-black frames at an act change, caused by a brightness dip into a lens bore. The fix was to crossfade the next scene in with a small scale settle. - A camera "jolt" on a magnetic close. Real contact is dead still. - An end-card push that carried the disclaimer to 31 px from the edge. The fix was to counter-scale the disclaimer. - A label clamp that fixed one safe-area bug and caused a collision two seconds later. - Hidden causes behind symptoms: effect gains solved against near-silent music, and a sample with a 33 ms lead-in that landed every hit 2 frames late. ## What the client said, in order 1. "The opening animation, I feel like you didn't nail it… research how they did it." This led to the reference study and the measured curve. 2. "Especially nail the folding animation." This led to screen continuity, the frost, stronger motion blur, dead-still contact and the fold critics. 3. "I feel like you didn't nail the blur." This led to the sharp wallpaper, the frost veil, the frosted cards and a longer shutter. 4. "The music is a bit off… use better music and it should be syncing." This replaced the AI score with human-made library tracks edited to the cut. 5. "Some of the sound effects are not good at all," "the whoosh… too loud". This cut the effects from 14 to 9 real recordings, with quiet air passes on real transitions only. 6. "Small sound effects are missing for all the little things… music −10 %." This added micro sounds on every small mechanical move, each at or below the music's local peak. See `audio.md` for the resulting sound rules.business-motion-film/references/case-study-home-display.md6.5 KB4fd734d051edupstream ↗
# Case study: a 20 s smart-home display spec commercial (Three.js-led, light and dark stages) An unofficial concept ad for a smart-home display, the day after it was announced, made in the brand's visual language. It is local only, and the end card carries an "unofficial concept, not affiliated" disclaimer. Every frame is code: seven Three.js canvases on one HyperFrames seek clock, with the type drawn on canvas above them. The device, its screen UI, the kitchen and every texture are procedural. There are no brand logos, no brand footage and no people. The critics said SHIP at v2. The client then asked for changes over seven more versions. **Almost all of them were about sound and pacing**, which is the main lesson of this film. ## Structure (as delivered, 18.5 s) | Time | Shot | What carries it | |---|---|---| | 0–1.5 | the device at night between the words of the title lockup | frame 0 is the finished lockup (the client's pick), with a slow push | | 1.5–4.5 | four colours side-on on a light stage, tilting one per 8th note | the brand's measured hinge curve; a highlight runs down the row | | 4.5–7.5 | monogram glass tokens drop through the cover glass; one person's Home Screen becomes another's | "the screen becomes a different person's in one continuous move" | | 7.5–10.5 | a section cut of the speaker: woofer plus two force-cancelling radiators in mirror | the parts are driven by the music's low band, so the speaker plays the score | | 10.5–14 | golden-hour kitchen: a touch dims the real pendants, the voice assistant closes the real blinds | the room equals the UI state, frame for frame | | 14–16 | the slat light carries across the cut onto the wall-mounted model | band angle, pitch and colour matched across the cut | | 16–18.5 | the end card plays the opening's distance-adaptive feature backwards | bookend | ## Process - **Component Gauntlet:** a lab, then a critic, then a fix, then a new critic for the device, the UI, every shot, the type system and the music. That was about 60 critic rounds, plus blind tests: our rendered frames mixed with the brand film's stills, for a judge who had to say which was which. - **A 7-window wow gate:** the 20 s split into 7 windows of ≈ 2.9 s, each of which must be judged YES. PARTIAL fails. - **Deadline mode:** with 1.5 hours left the client asked for 1080 only. Every in-flight builder got a 20–25 minute hard stop: land what works and revert anything half-done. Then: integrate, render at 1080, one full-film critic, its top 5 fixes in parallel, and a verify critic. Components whose last round wasn't re-verified were listed in the honest limits. - **At the critics' SHIP (v2):** - 6 of 7 wow windows YES; the opening was PARTIAL (small and dark for its first second); - the film blind test at 18/20, the first time a judge was fooled; - strict frozen time 0.15 s, with 0 single-frame cell pops; - −29.9 LUFS. ## What the client caught after the critics said SHIP Each item was named from one frame or one listen. 1. **"So much empty space"** on the glanceable UI face: tall tiles with the title at the top, the number at the bottom and a dead band between them. Fix: shorter tiles, each number centred between its label and its bar, and a third row whose content matches the headline. The headline said "lights, locks, blinds", so a Locks row went in. 2. **Text over text.** A voice-assistant request faded out while the card under it faded in, so both were readable at once. Fix: the request's words and the orb leave while the frosted sheet behind them is still ≥ 93 % opaque. The critics had checked overlaps on settled frames, not mid-crossfade. 3. **A state label that lied.** The label read "Open" while the percentage counted down. Fix: Open → Closing → Closed. 4. **"Start the video with this frame."** The client didn't want the reveal from the dark; they wanted the finished lockup at frame 0. Then: "the first scene lingers, start from 1.5". Fix: time-remap the shot's own clock rather than rebuilding it (a window offset plus a rate), and shorten the film to 18.5 s. 5. **"The music doesn't match at all… unnecessary sudden cuts… too loud."** This was the music editor's cut: a dead stop before the drop and a spliced final hit. Two rounds of library edits and two rounds of generated scores followed. What finally worked was **a real track back-timed so its own ending lands on the end card, one continuous piece, every cut on its beat**, offered as seven versions to choose from by ear (`audio.md`). 6. **"Every minimal movement should also have a sound."** The film went from 12 effects to 28. 7. **"I don't like that sound"** on the blinds countdown (a textured wood rattle) and on the speaker opening (a swoosh). I removed them, and the client corrected me: **"I meant use clean ones instead of random."** Replace a disliked sound with a cleaner one from the same family; don't delete it. 8. **"The tilt sound should be louder. Maybe we can't even hear it."** The micro-sound rule ("peak at or below the local music peak") had capped the hero's mechanical moves at 2–4 dB under the music's peak. The fix was explicit caps, which put the tilts at about +8–10 dB in-band and the tokens at +7–8. 9. **"A unique sound for the spatial-audio moment… now reduce it a little."** A sub knock plus a stereo-widened swell that peaks as the title lands. It settled at about +4.7 dB in-band. ## Render plumbing this film added See `product-hero-realism.md` §5. In short: - devicePixelRatio flips inside a 4K render; - DOM type re-rasterises in one-frame steps, so the type moved to canvas; - fonts load lazily; - readiness has to be registered early; - a NaN made one black frame; - hard thresholds on continuous values cause one-frame flashes; - the fast-capture path dropped a WebGL canvas. ## Delivery hygiene - **Desktop names:** another session wrote a file with the same name to the Desktop at the same time. Give deliverables names unique to the project, and put each round in its own named folder that opens when it's ready. - **Git:** the main checkout was on another branch with someone else's uncommitted work. Use a `git worktree` for the film's branch, and scan the staged files for secrets and oversized media before every commit. - **API keys:** keys pasted into a chat stay in its history; ask for them to be rotated afterwards. An automated agent may (rightly) refuse to read a key file. Give the user one command to run in their own terminal, without the `!` prefix: in zsh a leading `!` negates the command and skips everything chained after it. Mind the plan's concurrency limit; this ElevenLabs plan allowed 2 requests at a time.business-motion-film/references/critic-prompts.md3.3 KB5b31face5c76upstream ↗
# Critic prompts Copy, fill the `<>` slots, and send each to a **fresh** agent. Keep your own reasoning out of the prompt. ## Component critic ``` You are an independent visual critic in a Gauntlet loop; you did not build this. Judge only rendered pixels. Artifact: <path to lab proof MP4> (<duration>s, 1920x1080, <fps>fps). A <Three.js / UI> component for a premium <industry> commercial (brand: <palette hex values>, <typeface>). It shows <one-sentence description of what it should communicate>. It replaces <previous version and why it was rejected, in the client's words>. Bar: premium SaaS launch films (realistic materials, confident camera, meaningful information, no wasted space). Method: extract frames with ffmpeg (a contact sheet every 0.1s, plus native frames at <key times>) into <scratch dir> and look at them. Crop into details: geometry joins, materials, label legibility, pin accuracy against the thing it points at, collisions, pops or holds. Report (under 450 words), also written to <path>: - Verdict: KEEP / REVISE / REJECT. - Defects ranked by severity, with timestamps and screen regions. - The 3 highest-value fixes, implementable in code. ``` ## Full-film critic ``` You are an independent, harsh film critic in a Gauntlet loop. You did NOT build this; judge rendered pixels and measurable audio, not intentions. Artifact: <path> (<duration>s, 1920x1080, 60fps, with audio). <What the business is and what the film sells.> Brand: <palette>, <type>. <If a previous cut exists:> Previous cut: <path>. Previous critic report: <path>. Verify which issues are fixed. Client's criteria (mandatory): <paste the client's own feedback verbatim>. Benchmark: premium SaaS launch films. <Reference files or the relevant entries from launch-film-notes.md.> Method: ffmpeg frames every 0.2s into timestamped contact sheets plus native frames at key moments, and dense 1/30s windows around every transition (<list times>). Compute frame-difference to find frozen stretches. Measure audio with ebur128 and check whether sound lands on visual events. Report (under 900 words), also written to <path>: 1. Per scene: time range, what's on screen, % empty frame, how many seconds it could lose, and ranked problems. 2. Places where 3D would add real information (not decoration). 3. Layout and transition defects with timestamps. 4. The top 6–8 changes ranked by impact, concrete and implementable. Be blunt; no padding. ``` ## Verification critic ``` You are an independent critic; you did NOT build this. Artifact: <new render>. The previous critic's report: <path>. Section timings moved by about <x>s (<new section map>). For every item in that report's top fixes and new defects, give FIXED / PARTLY / STILL PRESENT with timestamps. Then list any NEW defects: glitch frames, overlaps, clipped text, awkward transitional frames. Check the audio hit at <time>. Write the report (under 500 words) to <path>, ending with SHIP or ONE MORE PASS (at most 3 fixes). ``` ## Storyboard critic ``` Review this storyboard for a <length>s commercial for <business>. You did not write it. Check: does the chronology match how the real service works; is every claim provably true; does each beat have a distinct composition and business job; which beats are filler; is the CTA unmistakable on mute? Return a ranked list of problems and one concrete fix each. ```business-motion-film/references/gauntlet.md4.9 KB5df6b09f6ca4upstream ↗
# The Gauntlet loop Adapted from the Gauntlet Loop idea (somethingbig.ai/gauntlet-loop): separate building from evaluation, compare against an inspectable reference, fix the biggest gap, repeat. ## Rules 1. **Set the outcome and bar first.** Write the brief, name the references, and state what "done" means (see `quality-bar.md`). 2. **Split the work into independently improvable parts.** Each 3D scene, UI component, transition and the audio mix is a part. 3. **Builder ≠ judge.** Spawn a *fresh* critic (a new agent or model session) for every review. Give it only: - the actual artifact (MP4 path or frames), not screenshots you chose; - the brief and the client's own words; - reference descriptions or files; - the previous critic's report (for verification rounds). Never give the critic your reasoning or a list of what you think you fixed; it must find things itself. 4. **Critics judge pixels and measurements**, not intentions. They extract frames themselves (contact sheets every 0.1–0.25s plus dense 1/30s windows around every transition) and compute frame-difference and loudness. 5. **Fix the largest meaningful gap first.** Don't average away a weak opening with strengths elsewhere. 6. **Verify, don't assume.** The next round uses a *new* critic that checks each prior item as FIXED / PARTLY / STILL PRESENT and looks for regressions. Fixes often create new defects, such as a covered element losing its z-order or a moved object now crossing a headline. 7. **Stop** at the quality bar, at diminishing returns (remaining items are sub-frame or cosmetic), or when the user says stop. Not after an arbitrary number of rounds. 8. **Keep a ledger.** One table: round → artifact → critic's top findings → changes made → measured result. It becomes the proof of craft for the client, and training material for the next project. ## Round types | Round | Artifact | Critic focus | |---|---|---| | Storyboard | Table + rough frames | Chronology, business logic, truthfulness, variety of compositions | | Asset | Generated stills/clips | Geometry fidelity (straight lines, believable joins), exposure, artifacts, fit to brief | | Component | Isolated lab render (stills + 3–4s motion proof) | Material realism, camera, readability, collisions, pops, what information it adds | | Full film | Whole render | Holds, empty space, transitions, text collisions, pacing, business clarity, audio | | Verification | New render + previous report | Item-by-item status + new regressions + SHIP / ONE MORE PASS | ## What critics caught in practice (real examples) - A one-frame "pop" when a 3D shingle layer started lifting: an ease-out started at full velocity, and individual courses separated in a single frame. Fix: smootherstep with tiny stagger. - Wipes that spliced two titles into one word ("Sur|ctions."). Fix: the outgoing title leaves before the wipe, and the incoming title enters after it. - A camera path that went near top-down, turning a house into a flat slab with a visibly repeating texture. Fix: limit pitch to 35–55° and keep a minimum distance. - A report card that went blank for 0.2s between states. Fix: crossfade contents in place. - Text flying through other text during a move. Fix: fade out, then fade in at the destination. - A retained heading misaligned by ~12px because its parent had a slow scale push. Fix: animate the overlay with the same push. - A photo labelled with the wrong caption, because one tween targeted both caption variants. - Music that faded out before the logo landed, and sound effects that were too loud and harsh. - Around 10 of 36 seconds effectively frozen. None of the builder's checks flagged it as a problem. ## Deadline mode When the client compresses the schedule (e.g. "we need 1080 and have 1.5 hours"): 1. Give every in-flight builder a hard stop of 20–25 minutes. They land what works and revert anything half-done. 2. Integrate, then render at the delivery resolution only. 3. Run one full-film critic, then its top fixes in parallel. 4. Run a verify critic and deliver. 5. List every component whose last round wasn't re-verified in the honest limits. ## SHIP is not the end On a 20 s spec ad the critics said SHIP at v2, with 6 of 7 wow windows YES. The client then found nine more things in seven rounds: empty space in a UI face, text over text mid-crossfade, a lying state label, the opening frame, a lingering first shot, music that "doesn't match", and three rounds of effect levels. Critics measure what they're asked to measure. Add those checks to the critic prompts (`quality-bar.md`, "Checks critics missed"). Plan for the client's own pass: deliver versions in a named folder, make each change by timestamp, and turn every round around in minutes. ## Measured progress example Frozen time (frame difference below threshold) went from ~10s/36s → 6.5s/31.3s → 1.1s/30.1s → 0.7s/28.6s across four rounds. Loudness range went from 0.6 LU (flat wall of sound) to 3–4 LU.business-motion-film/references/launch-film-notes.md14 KBe950c4a8624eupstream ↗
# Launch film notes (28 references) The written learnings from 28 professionally produced launch films listed in Skale's public portfolio (skale.solutions/portfolio). The films belong to their owners and are **not** included here; links point to the original posts so you can watch them. Timestamps refer to those original posts. View and reach numbers in the portfolio are studio claims, not verified results. Each entry lists what the film is, one studied moment, the mechanism behind it, and how that mechanism transfers to a business video. The roofing examples come from the case study in `case-study-alder.md`. Strongest motion-first references: **Contra 24, Wonder 18/19, Bolt 25, Work Louder 22, Replit 11–13, Contra 23, Extend 28, Poke 17, AsideHQ 05**. --- **01 · TypeSafe** ([post](https://x.com/completeskeptic/status/2099925682726002904)): founder speech intercut with editorial flow diagrams, pink/white/black type, code comparisons and voxel forms. - **Studied moment:** around 24s, an anchored node and connector stay put while the text changes and a contrasting label arrives. - **Transfer:** labels can organise an explanation. Keep one visual artifact on screen while the next proposition arrives. Long speaker sections don't suit short ads. **02 · T:0 finance agent** (portfolio label "Airwallex") ([post](https://x.com/lancectk/status/2094830760410984550)): white/black finance typography, green annotation marks, connected inputs, agent UI. - **Studied moment:** around 16s, words rearrange into "the pre-AI era", then a green hand-drawn strike lands. - **Transfer:** fast semantic type changes plus one accent beat. **03 · Cognition, Devin Voice** ([post](https://x.com/cognition/status/2098142686486356185)): dialogue, a plush character, phone, documentary warmth, thin charts over footage. - **Studied moment:** around 21s, a chart stays registered over the character while a caption advances. - **Transfer:** restrained overlays explain a concept. Not a pacing model for motion-first ads. **04 · Poke × Cognition** ([post](https://x.com/cognition/status/2080311229256540194)): founders, filmed phone UI, warm vignettes, dimensional phone cards, acquisition close. - **Transfer:** keep the identity of the real business action visible. Holds work when the thing being held is the product action. **05 · AsideHQ** ([post](https://x.com/hyojun_at/status/2069497198879048131)): a dark "friction" premise gives way to a bright browser workspace doing concrete tasks. - **Studied moment:** around 18s, a cursor crosses dock icons and the selected app surface expands upward like a fan into the next scene. - **Transfer:** **an action causes the transition.** Use it for a "request" button expanding into the next state. **06 · Browserbase** ([post](https://x.com/pk_iv/status/2041518621290266632)): playful bitmap nature, coloured markers, prompt UI, red cursor labels, staged browser tasks. - **Studied moment:** around 12s, the same page reveals lower cards while the camera pulls back and the cursor keeps tracking its target. - **Transfer:** overlap nested UI with camera motion and keep the same working surface. **07 · Poke, drones / flamethrower** ([post](https://x.com/interaction/status/2034713714415608123)): long portrait testimonials and demo, then a travel/phone montage. - **Studied moment:** around 850s, the phone stays centred and sharp while the background streaks past. - **Transfer:** keep the subject stable through moving surroundings. **08 · ListenLabs** ([post](https://x.com/itsalfredw/status/2011469284749594774)): a cream/blue/mustard system; circles turn edge-on and repeat in depth; type expands beyond the frame. - **Studied moment:** around 17s, a capsule outline draws around a placeholder that swaps to a typed question, with a stable arrow target. - **Transfer:** custom contours and text-state changes can animate together, and extreme scale contrast gives type energy. **09 · Buds** ([post](https://x.com/budapp/status/2046605073741119800)): a warm yellow phone/browser sequence; prompt → documents → results; a desktop agent demo. - **Studied moment:** around 46s, a result moves up, its text follows, and the next result arrives from below, all inside one workspace. - **Transfer:** change content inside a persistent workspace instead of a sequence of unrelated slides. **10 · Taste** ([post](https://x.com/thaiscbranco_/status/2066912871649574945)): founder footage, black design systems, perspective galleries, artifact rings, network scenes. - **Studied moment:** around 16.3s, a diagonal perspective wall of large panels becomes an elliptical belt of small screenshots orbiting a centred sentence, which keeps adding words. - **Transfer:** a compact orbit of business artifacts (request, photos, appointment, quote) around one stable sentence. It stays smooth because the centre of attention and the cards' track are shared. **11 · Replit Slides** ([post](https://x.com/Replit/status/2049160182488576245)): cream background, black type, orange emphasis, oversized controls, named cursor. - **Studied moment:** around 16.4s, green dots resolve into a card that grows while a colour wash reveals its regions in stages, then settles larger. - **Transfer:** **materialise results.** The shell appears first, then status, date and action fill in order, and the completed state stays readable. **12 · Replit Canvas** ([post](https://x.com/Replit/status/2060097656207413613)): prompt → design directions → canvas → editing → publishing. - **Studied moment:** around 21.7s, the camera travels down a column of artboards; neighbours blur while the selected column stays sharp, and new content appears before the old leaves. - **Transfer:** one continuous canvas from request → schedule → confirmation, with one selected card staying crisp and a single camera transform over nested UI. **13 · Replit Parallel Agents** ([post](https://x.com/Replit/status/2053891504989753817)): red pill words, task cards, branching lines, one main app receiving updates. - **Studied moment:** around 42–45s, four endpoints travel along thin rails, the headline swaps left-to-right, and the rails persist across the swap. - **Transfer:** a persistent path connects the steps of a process, so words change without resetting the scene. Show merging only when the real process merges. **14 · Adaline** ([post](https://x.com/arshdilbagi/status/2065826083224834345)): code, founder, green/cream diagrams, orbital paths, radial networks. - **Studied moment:** around 25.8s, a globe and orbits dissolve into a radial network, and the next phrase forms in its centre. - **Transfer:** a small consistent shape family (circles, thin lines, compact labels) can carry complex ideas quietly. Avoid low-contrast pale elements on the key message. **15 · Pilot Protocol** ([post](https://x.com/razvanr/status/2081756497814720657)): founders with huge changing text overlays, stark black/white graphics, lime keywords. - **Studied moment:** around 52.5s, black tiles multiply over the founder, then expose a new environment. - **Transfer:** a tile wipe makes a good one-time chapter break, not a main transition language. **16 · Bevel** ([post](https://x.com/greynguyen/status/2048799663638171746)): soft white space, lavender/mint light, isolated mobile screens, curved artifact rows. - **Studied moment:** around 16s, screens arrive upward in a stack; each enters soft-focused, aligns and occludes the previous one. - **Transfer:** stack three business cards, then collapse them to the selected one. Overlapping timing plus decreasing motion as they align reads as smooth. Keep the stack shallow so the text stays large. **17 · Poke, seven-video campaign** ([post](https://x.com/interaction/status/2062575428213285352)): a portrait UI story (messages → calendar picker → choices → results → reminder). - **Studied moment:** around 3–6s, an avatar pulls back into the phone; a request bubble stays readable while the background dims and a calendar sheet rises from below. - **Transfer:** **request → picker → selected time → confirmation** is the most directly usable sequence for service businesses. Keep the request as the anchor, and make every scene show an action and its visible result. **18 · Wonder (first video)** ([post](https://x.com/aibek_design/status/1957856991608512898)): a purple cursor, selection outlines, canvas framing, generated screens, component extraction. - **Studied moment:** around 11–13s, the camera pulls back from a selected frame while skeleton regions appear in neighbouring artboards; the cursor keeps a meaningful target. - **Transfer:** couple the global camera move with local UI motion, keep selection borders attached to their objects, and make sure the final state is readable after the energetic reveal. **19 · Wonder (second video)** ([post](https://x.com/usewonder/status/2044099145997402272)): a dimensional cursor types a title, then prompt → moodboard → website → variants. - **Studied moment:** around 2.8–3.8s, the selected title barely moves, then accelerates through the camera with stretched, blurred edges while white space clears for the next prompt. - **Transfer:** a photo selection frame expands into the detailed inspection, then becomes the quote panel. The cursor or selection marker remains the actor. **20 · MadeThis** ([post](https://x.com/jehovahscript/status/2082868450847031354)): a historical intro, then a blue/pastel SaaS system; a typed business prompt creates brand, shop, payments and marketing. - **Studied moment:** around 20s, a tilted wheel of business labels advances; the active line is bright, neighbours recede, and the background hue flows. - **Transfer:** a carousel of service steps (Inspect / Measure / Quote / Schedule) with one readable active item. **21 · AgentArcade** ([post](https://x.com/AgentArcade/status/2077071110777029066)): a dark neon agent UI, chart cards, fanned pills, ranked rows, radar diagrams. - **Studied moment:** around 12.2s, an upright panel folds into a luminous strip and the next card rises from below. - **Transfer:** a perspective fold turns an estimate into a scheduled-job card. Use stronger contrast than the source, because near-black transition frames hide the story on phones. **22 · Work Louder** ([post](https://x.com/MarkKnd/status/2022261281114468759)): keyboard macro sweeps, a suspended product, glowing key waves, then exploded physical layers with a flowing cable. Almost no copy. - **Studied moment:** around 12.9–13.9s, parallax keyboards **hard-cut** to three exploded layers that keep drifting. It stays cohesive because geometry, palette and motion continue. - **Transfer:** **the best physical-product reference.** Separate a product into its real layers (e.g. shingles / underlayment / deck / rafters), with disciplined camera and materials. **23 · Contra (Indy)** ([post](https://x.com/contraben/status/1952725885146075406)): "Job boards are dead"; a black wall of posts; mascot; tiled grid; scanning counter; results. - **Studied moment:** around 7s, a wall of posts races upward, cuts to a mascot close-up, then pulls back into a clean grid. - **Transfer:** **information overload → one clear actor**, then a scan that visibly produces results. A wall of inspection photos resolves to one selected roof, then to an estimate or booking. **24 · Contra (Creative Arena)** ([post](https://x.com/contraben/status/1968717444685394134)): a big "AI" premise, a constellation of creative panels and tool logos, a dive into the product, a selected vote, results. - **Studied moment:** around 9.3–10.3s, the centred claim and logos scale through the camera with stretched letters while the destination website is already visible underneath, then settle. - **Transfer:** **the model foreground fly-through.** The business promise flies past to reveal the product or measured object. Show two options, select one, and let the selected panel become the next scene. **25 · Bolt.new** ([post](https://x.com/boltdotnew/status/2039024467846860972)): giant kinetic problem/solution type, blue light on black, code panels, prompt → site, merge → confirmation. - **Studied moment:** around 19s, the brand letterform scales up until its hole fills the frame and becomes the next scene. - **Transfer:** use a brand shape (e.g. a roofline) as a portal; action → button → "ready/booked" status with truthful wording. Strike-through text replacement is another strong move. **26 · Tembo** ([post](https://x.com/benjaminakar/status/1985754174911627546)): minimal black-on-white; automation setup screens; result panels; template cards. - **Studied moment:** around 40s, one card moves to centre, then its copies fan diagonally as the camera pulls back. - **Transfer:** one inspection becomes a fan of options or completed projects, with the selected front card preserved. The overall pace is too calm for short ads. **27 · Conduit** ([post](https://x.com/0xpunnk/status/2062598516544061537)): a hospitality image stack, a message constellation, founder footage, action tags, a gradient brand reveal, an ops demo. - **Studied moment:** around 28.3s, a soft coloured cloud takes over the frame as the logo settles. - **Transfer:** one colour-field breath for the brand handoff. Its clear input → work → result business vocabulary is the real lesson. **28 · Extend** ([post](https://x.com/kushalbyatnal/status/2059278322287214750)): a big statistic, an orbiting document sphere, benchmark charts, an isometric document with coloured semantic regions. - **Studied moment:** around 64–67s, a perspective document keeps its registered colour regions while a scan completes; the regions lift as planes, then category cards reveal diagonally. - **Transfer:** **registered decomposition.** The source object stays recognisable while its parts become structured outputs (roof plan → measured planes → scope items). --- ## Cross-film takeaways for business ads - The strongest films never cut to an unrelated layout without a carried object, matched direction or pre-positioned destination. - Physical products get material realism and exploded structure (Work Louder); software gets cursor/selection causality (Wonder, Replit); services get request → choice → confirmation (Poke). - Long founder/talking-head grammar and slow historical intros don't fit a 15–30s business ad. - Near-black transition frames and pale low-contrast text on key messages hurt phone viewing.business-motion-film/references/motion-grammar.md4.8 KBe5a0f05cb47bupstream ↗
# Motion grammar (distilled from 28 professional launch films) These principles came from studying 28 professionally produced SaaS launch films (48 minutes, 81,647 frames numerically analysed, plus sampled visual study of every film and dense frame-by-frame study of key transitions). Per-film notes are in `launch-film-notes.md`. ## The six rules 1. **The foreground becomes the transition.** A headline, logo or shape scales through the camera, with directional blur and stretched edges, while the next scene is already in place underneath. There's no empty gap between scenes. (Contra "Creative Arena", Wonder, Bolt.) 2. **One persistent actor.** A selected photo, cursor, card, rail or measured shape keeps its identity across shots. Unrelated reveals in sequence feel like a slideshow however well each one is animated. (Wonder's cursor, Replit's rails, Poke's request bubble.) 3. **Density comes from hierarchy.** One primary camera/object move, supporting staggers on cards and labels, and fine tertiary detail (edge sweeps, status ticks), all overlapping. Don't start and stop the whole frame on every beat. (Taste's artifact ring, Wonder's canvas.) 4. **Change speed; don't drift at a constant rate.** Readable target → fast exit → slowing arrival. Motion eases into a landing so the viewer can read, then leaves quickly. 5. **Smooth doesn't mean dissolves only.** Hard cuts work when subject, scale, direction and material language match and motion continues after the cut. (Work Louder cuts from parallax keyboards straight into an exploded view.) 6. **Show the business purpose as cause → effect.** Scans produce findings, prompts produce artifacts, buttons produce states. Every animated action should visibly produce a result. (Contra's scanning counter → results; Bolt's merge button → "ready".) ## Composition rules that follow - The lead subject fills 60–85% of the usable frame in feature beats. Empty space reads as cheap to buyers. - Vary composition types: physical macro, wide dimensional, overhead/plan, UI close-up, typographic impact. Don't repeat "heading above three cards". - Keep 1–2 reading targets at a time. Dense doesn't mean everything competes. - Type is part of the choreography: oversized words enter on opposing axes, mask or reveal the next scene, and never sit still over busy footage without a backing. - Hand objects across scenes at exact pixel coordinates. Pre-position the destination so the carried object lands on it. - Keep one wipe direction and one title position through a sequence of similar shots. The outgoing title leaves before the wipe; the incoming title enters after it. ## Mechanism catalog (reusable moves) | Mechanism | What it does | Seen in | |---|---|---| | Foreground fly-through | Claim/logo scales past the camera, revealing a pre-positioned next scene | Contra 24, Wonder 19, Bolt 25 | | Logo-as-portal | The brand shape grows until its hole becomes the next frame | Bolt 25 | | Selection → expansion | A selected frame expands into the detailed workspace for the next beat | Wonder 18/19, AsideHQ 05 | | Materializing result | A shell appears first, then its internal regions fill in order | Replit Slides 11 | | Persistent rails | Lines/paths survive a text swap and connect two states | Replit 13 | | Continuous canvas | The camera moves over one spatial layout; the selected item stays sharp while neighbours blur | Replit Canvas 12 | | Artifact ring | Many small artifacts orbit a stable centre sentence | Taste 10 | | Card stack / fan | Screens arrive and occlude; one card becomes a collection, or a collection resolves to one | Bevel 16, Tembo 26 | | Perspective fold | A panel folds to a luminous strip; the next panel rises out of it | AgentArcade 21 | | Exploded layers | A product separates into its physical layers, then recombines | Work Louder 22 | | Registered decomposition | The source artifact stays recognisable while regions lift into categories | Extend 28 | | Wall → one actor | Information overload collapses to a single clear subject | Contra 23 | | Carousel emphasis | A rotating list with one bright active item and receding neighbours | MadeThis 20 | | Colour-field takeover | A soft gradient cloud fills the frame for a brand handoff (use once, not as the main vocabulary) | Conduit 27 | | Tile wipe | A grid of tiles as a chapter break (fragmented, use sparingly) | Pilot Protocol 15 | | Request → picker → confirmation | A message bubble anchors while a sheet rises, a choice is made, and the state confirms | Poke 17 | ## Pacing numbers that worked (30s business film) - About 12–15 compositions; beats of 1.4–3.5s; no motionless stretch over ~0.6s. - Hook: frame 0 must be a finished composition, not a half-entered word. - Reading holds only on the few frames that carry the business message (report, CTA); even there, a slow 3–5% push keeps it alive. - End card: about 1.5–2s readable, with gentle drift.business-motion-film/references/product-hero-realism.md9.1 KB0c5737d6edf5upstream ↗
# Product hero realism: making a Three.js device move like the real one This comes from building a 40 s spec commercial for a foldable phone. The client's main note was: "especially nail the folding animation… it should look realistic". Everything here is measured, not guessed. ## 1. Study the reference motion frame by frame - Download the brand's own film for **study only**; it never goes into the output. Tile the hero motion at 10 and 30 fps (`ffmpeg -ss A -t D -i ref.mp4 -vf "fps=30,scale=400:-2,tile=7x7"`). Tile your render at the same fps and crop, and compare side by side. - **Measure the curve.** Estimate the angle per frame from the projected width of the moving part and write keys (seconds after start → degrees). A real foldable unfold ran 0→180° in about 0.8 s, with these keys: `0→0, .08→20, .15→58, .23→82, .31→98, .39→116, .49→140, .59→159, .69→172, .80→180`. It is fastest at 30–60°, slows (dwells) near edge-on and lands softly. That is **not** a smoothstep. - Interpolate the keys with a **monotone cubic (Fritsch–Carlson) with zero end slopes**. Blending keys with a smoothstep "to soften it" moved the dwell and created a mid-swing hiccup (7.8 → 2.6 → 5.4°/frame), which a critic caught from a per-frame CSV. Have a critic plot °/frame: it must peak once and then dwell, with no second hiccup. - **Lock the camera** during the motion. A reframe that stops on the contact frame reads as a camera bump. Re-centre the *object* during the swing (the real footage does this), and move any camera reframe to after the landing. ## 2. The screens are half of the realism The motion matched early on, but it still read as CG until the screen behaviour matched the reference: - **Before the move:** the outer display shows the *home* screen, reached via a quick unlock. Stagger the unlock: the lock layer lifts and fades out, *then* the home screen scales in from 1.06. A simple crossfade gave a doubled "9:41", which counts as a text collision. - **During the move:** the outer display **blurs but stays lit** on the moving cover and only sleeps near edge-on. Our first version slept at the lift; that made a 74 % luma drop and a black hole on a black stage. - **Continuity:** the big display's right half shows *exactly* the layout the small display showed, pixel for pixel. Build the small display's texture as a crop of the big one, so the match holds by construction. - **The frost:** the newly revealed half keeps a sharp, continuous wallpaper under a white frost veil (about 0.3). The new widgets arrive as frosted, whitened, translucent cards (blur about 40 px, 0.6 opacity, 0.97 scale). These resolve to sharp in about 0.35 s *after* the landing. Blurring the whole half, wallpaper included, read as a smeared photo. Pre-render about 6 blur levels and crossfade between them, so the frost is cheap per frame. ## 3. Motion blur that holds up - Use deterministic accumulation. Render N sub-frames inside the shutter (`t + ((i+½)/N − ½)·shutter`), average them into a half-float target, and keep every frame a pure function of t. - **Sub-frames:** 10 left a visible comb on a fast edge (a 3 px ripple of ±20 luma). 24 over the fastest 0.5 s cleaned it up. - **Shutter:** 180° looked too crisp next to the real footage. About 300° during the fastest swing, easing back to 180°, matched it. - **Switching:** ramp the shutter from 0 to its full angle over about 6 frames at the window edges, so it never switches on in one frame. Check that no brightness step appears where blur turns on or off. - **Coverage:** every fast fold in the film needs it. One fold was moving 52 px/frame, another 35 px/frame, and both had no blur. ## 4. Light that tells you the angle - **Keep ≥60 % of the glass reflections through a fold.** Our builders cut 85 % to hide one-frame flashes, and that removed the main cue that the angle was changing. - **Fix flashes by widening the light, not dimming it.** A flat polished rail sweeps past a narrow light's mirror angle within one frame. Widen the strip (width 3 → 12 at low power) so the glint builds over about 6 frames. Find the culprit by switching lights off one at a time. - **Build a black cover to read on a black stage.** Place a broad soft box at the mirror position of the camera about the glass, and let it travel with the angle. - **Keep magnetic contact dead still.** The real close is pixel-identical for about 17 frames after contact. A camera "jolt" (about 22 px) and a 1.7° rebound read as fake; the fix was no jolt and a rebound of about 0.27°. ## 5. The render plumbing - **Handoffs:** use shared functions (`applyTurn(u)`, `applyPoseP()`) so both sides of a cut compute identical poses. Check each cut with a mean frame difference against its neighbours. - **The environment map:** three r181 `PMREMGenerator.fromScene` ignores the render target you pass and leaks a new one every call. Keep one persistent target, copy each result into it and dispose of the fresh one. Feather the soft boxes, or the reflections have hard edges. - **Supersampling:** render at a 1.5× pixel ratio and let the browser downsample; size the depth-of-field target to the drawing buffer. - **Fonts:** headless Chrome renders `-apple-system` as Times. Bind the family names with `@font-face` to a local font file that sits next to the CSS (lint rules may reject `../`), and don't redistribute system fonts. - **Shared-file edits:** while several builders work in parallel, edit shared files through a `.next.js` copy, run `node --check`, then `mv` it into place. One mid-edit syntax error blanked every canvas. - **Stage colour:** measure the stage colour through the capture→encode path. CSS `#f5f5f7` decoded as `#f3f5f5`, so author the stage colour to what should come out. - **Blind tests on encoded frames only.** Lossless PNG captures of our stage sat at (245,243,246) while the reference film decoded to (243,243,245): a perfect tell that has nothing to do with realism. Pull your frames from the rendered MP4, run both sets through the same encode, and hide the key. **More gotchas, from a 20 s spec ad (HyperFrames 0.8, 4K = the 1080 composition at a device pixel ratio of 2):** - **Never read `window.devicePixelRatio` per frame.** Inside a 4K render it reads 2 only on frames 0, 30, 60… and 1 in between, while the pixels stay true 4K. Anything sized from it per frame (mesh sharpness, line widths) flickers. Freeze it at load. - **Render with `--experimental-fast-capture=false`.** The fast path dropped a WebGL canvas at 4K. - **Type under a slow transform steps.** DOM text under a slowly changing CSS transform or font size re-rasterises in visible one-frame steps (≈ 10 steps in 2 s at a 12 %/s push). Draw each line on a 2× canvas: one `fillText` of the whole line, with each word revealed through a clip so the per-word rise keeps the natural kerning. Keep the DOM line at opacity 0 for layout and overlap checks. A transparent fill gets flagged as unpainted text. - **Fonts load lazily.** A line laid out before its weight loaded kept the fallback font's metrics. Preload every weight (`document.fonts.load`) before mounting the overlays. - **Hidden segments measure 0 px.** Overlays mounted inside `display:none` segments measured zero-width text and overlapped. Lay every segment out (visible but hidden) while the overlays mount, then hide them. - **Register readiness early.** Late registration of the render's ready promise captured blank first frames. Do it in a classic `<head>` script, before any module. - **`will-change` broke seek determinism.** Remove it. - **One NaN, one black frame.** A ripple shader evaluated past its range produced a single black frame that existed in every render for two rounds. Scan every render for single near-black frames. - **Hard thresholds on continuous values flash.** A bulb icon switched its ink at `light > 0.3`, and the animation landed at 0.29999999, so the bulb was white for exactly one frame. Make every threshold on a continuous input a crossfade. - **Comments can swallow code.** A trailing `//` comment on a long one-line statement swallowed the next uniform definitions. `node --check` passed, and the scene threw at load. Render a frame after every edit to a shared file. - **A whole-frame "pop" flag may be intended.** A screen blooming from dark to light and a woofer making a full stroke both tripped the single-frame pop detector. Before you "fix" a flag, look at the frame-to-frame difference over ±5 frames: a ramp (0.25 → 1.08 → 0.36) is motion, while a lone spike is a bug. - **Re-time a shot instead of rebuilding it.** Give each segment window a local clock offset, and let a scene remap its own clock (`t → t0 + rate·t`). Opening on a later, finished frame and trimming a lingering first shot were both one-line changes. ## 6. How the review loop found these - The critic that mattered most compared the render against the reference **frame by frame** and wrote a **per-frame CSV**. - The director's own first reading of the reference was wrong once: whether the outer display sleeps at the lift. A second, fresh fold critic caught it. - Always verify the fixes in the next round; don't trust the builder's own numbers. One builder measured a glint at 1.11× its neighbours; the critic, measuring in the film, found 2.1×.business-motion-film/references/quality-bar.md3.9 KBf168f68ed12eupstream ↗
# Quality bar A cut ships only when all of these hold. Critics check them; scripts measure the numeric ones. ## Measured | Check | Target | How | |---|---|---| | Frozen time | Total stretches with frame-diff below threshold ≤ ~1s per 30s; no single hold > 0.6s except the final CTA | `scripts/frozen-time.sh` | | Loudness | −14 LUFS for punchy launch-style pieces; about −16 LUFS for calm pieces; true peak ≤ −1 dBFS | `scripts/loudness.sh` | | Dynamics | Energetic scores: loudness range ≥ ~3 LU (a flat wall of sound reads as stock). Calm/chill scores: ~1.5–3 LU is fine, as long as the track doesn't die before the ending | `scripts/loudness.sh` | | Effects vs music | ~+4 dB lift within each effect's own frequency band over the music-only mix; never inaudible, never poking out (clicks/ticks spiking 10+ dB at 2–8 kHz read as annoying) | Band-passed 50ms peak vs music-only render | | Text contrast | WCAG AA on all settled text (transitional fades excepted) | Framework check / manual sampling | | Brand colour | Sampled backgrounds match brand tokens (tone mapping in 3D shifts colours, so measure) | Crop + average pixel | | Determinism | Same frame rendered twice from different seek orders gives identical pixels | Seek-consistency test | ## Visual (critic-judged) - **No empty frames**, e.g. a blank colour band before a title lands, or a lone photo on an empty field. - **Frame 0 is a finished composition.** No half-entered word at the first frame. - **No text collisions,** including mid-transition: one title over another, text flying through a heading, titles spliced by a wipe. - **Equal spacing** in lists and rows (identical row height, shared left edges, separators). Clients notice this immediately. - **Aligned left edges** across headline, image and supporting rows. - **Lead subject fills 60–85%** of the frame in feature beats. - **3D reads as designed:** no toy-like gaps, floating parts, visible texture tiling, top-down "slab" angles, or pops. - **Pins and callouts** sit on the thing they name, and each photo's label matches its content. - **Transitions** keep a carried object or matched direction. No unrelated slide-in after unrelated slide-in. - **Brightness:** home-services and consumer buyers rejected dark, moody openers. Keep imagery bright and clear unless the brand demands otherwise. ## Business (critic-judged, on mute) - By the end, a first-time viewer can say what the business does and the one next action. - Every claim is true or clearly labelled as a concept. - The CTA is readable at phone size for at least ~1.5s. ## Common rejection reasons from real clients - "Too basic: image, then video, then text." - "So much space is being wasted." - "Some parts linger too long." - "Not spaced equally." - "It's not elite; go harder." - "The first image looks too dark." - "The music doesn't match." - "The sound effects feel unprofessional and jarring." - "So much empty space" (a glanceable UI with a dead band between the title and the number). - "Start the video with this frame" / "the first scene lingers" (open on the finished hero composition, not a reveal). - "Unnecessary sudden cuts" in the music (dead stops and splices), and "too loud" at −30 LUFS. - "Maybe we can't even hear it" (the hero's mechanical moves held under the music by a micro-sound cap). - "I meant use clean ones instead of random" (a disliked effect should be replaced, not deleted). ## Checks critics missed and clients caught - **Text over text mid-crossfade.** Check overlaps on every frame of every crossfade, not only on settled frames. - **State labels that contradict the values**, e.g. "Open" while the number counts down. - **Glanceable layouts:** no dead bands; each big number centred between its label and its bar; every tile earns its space. Content that echoes the headline (a "Locks" row under "lights, locks, blinds") fills the space with meaning. - **The hero moments are audible** on a laptop at normal volume, with the music on.business-motion-film/references/three-js-patterns.md4.3 KB5bf1768d23c9upstream ↗
# Three.js patterns for commercial motion Use 3D where it **explains** something physical or spatial. Everything else (type, reports, forms) stays crisp HTML/SVG. ## Render contract (deterministic, seekable) ```js export async function createScene(canvas, opts = {}) { const renderer = new THREE.WebGLRenderer({ canvas, alpha: true, antialias: true, preserveDrawingBuffer: true }); renderer.setPixelRatio(1); renderer.setSize(1920, 1080, false); renderer.outputColorSpace = THREE.SRGBColorSpace; renderer.toneMapping = THREE.ACESFilmicToneMapping; // ...build everything once... function pose(t) { /* set every transform from t only */ } function render(t) { pose(t); renderer.render(scene, camera); } function anchors(t) { pose(t); return points.map(p => project(p)); } // screen coords for HTML overlays await renderer.compileAsync(scene, camera); render(0); return { render, anchors, dispose }; } ``` - The root timeline calls `render(localTime)` on every seek and update. No RAF loop, timers or physics history. - Load all textures before resolving, and have the renderer wait for fonts and every scene before capture. - Retime by mapping film time → local time (`(t - start) * speed`), never with a free-running clock. ## Patterns that worked **Exploded architectural layers.** Build the object in its own local frame (e.g. for a roof plane: x along the ridge, y out of the plane, z down the slope, via `matrix.makeBasis(u, n, v)`). Stack real layers (rafters → deck → underlay → shingle courses) as separate groups, then lift each along the plane normal with different amounts and a smootherstep ease (`x*x*x*(x*(x*6-15)+10)`). Use a tiny stagger (≤5ms per course). A cubic ease-out starts at full speed and makes a visible one-frame pop. After the lift, keep a slow "breathing" separation (+10–15%) so the hold isn't frozen. **Projected callouts.** `anchors(t)` returns screen points (`vector.project(camera)` → pixels) for points on each layer. HTML labels sit in an evenly spaced column with leader lines to the points; recompute every frame. Equal label spacing matters more than exact alignment to each anchor. **Photos pinned in context.** Fly the camera around the model; at the moment each documented spot is visible, a real photo card pops out of its pin with a leader line, then docks into a rail. On handoff, cards animate to the exact pixel rectangles of the next scene's layout. Keep cards opaque with a shadow that fades during handoff. Pin on the side facing the camera, and make each label match its photo. **Animated option patches.** Pixel-space orthographic camera (`OrthographicCamera(-960, 960, 540, -540)`) so each 3D patch sits exactly in its HTML card. Give each option a distinct end state: e.g. maintain = a scan line that stays parked, repair = one visibly new tile, replace = a new field with an edge trim. Sequence actions so only one piece is airborne at a time, and keep moving parts inside their card. **Physical logo mark.** Bevelled parts assemble into the mark, then the same canvas moves (CSS transform) into the final lockup, so the mark visibly survives into the end card. ## Realism checklist (what critics flagged) - Background colours shift under ACES tone mapping. Measure the rendered pixel and compensate with an HDR background (e.g. multiply the linear RGB by 1.1–1.9 per channel) until it matches the brand token. Fog uses the same colour. - Architectural-model ("maquette") styling reads premium and avoids uncanny fake photorealism: cream walls, real roofing material textures, a model base board. - Toy giveaways: gutters floating off the fascia, square downspouts that read as posts, flat black glass (raise the colour and add environment reflection), open ridge notches (add a ridge cap), plinth-like flashing (make it a flush sloped apron). - Studio lighting: a soft key from the side you want lit, low shadow radius so lifted layers cast readable shadows, and a warm/cool rim. - Camera: pitch 35–55°, keep a minimum distance (never top-down on large textured planes, where tiling becomes visible), no roll, pull back rather than end tight on a slab. - Depth of field: a subtle lens pass helps macro moments. Focus on the subject distance. ## Lab first Build each scene in `templates/component-lab.html`, render stills at meaningful times plus a 3–4s proof, and get a component critic's KEEP before integrating.business-motion-film/templates/component-lab.html2.2 KB09c81ecb1e54upstream ↗
<!doctype html> <!-- Isolated lab for one Three.js component (HyperFrames-compatible). 1. Put your module next to this file as ./assets/my-scene.js exporting createScene(canvas) -> {render(t), anchors?(t)}. 2. Set DURATION. 3. Snapshot stills: npx hyperframes snapshot . --at 0,1.2,2.4 -o shots Motion proof: npx hyperframes render . --fps 60 --output proof.mp4 4. Send proof.mp4 to a fresh component critic (references/critic-prompts.md). --> <html><head><meta charset="utf-8"> <script src="https://cdn.jsdelivr.net/npm/gsap@3/dist/gsap.min.js"></script> <style>html,body{margin:0;width:1920px;height:1080px;overflow:hidden;background:#f4f0e5;font-family:system-ui,sans-serif} #lab{position:relative;width:1920px;height:1080px;overflow:hidden}canvas{position:absolute;inset:0;width:1920px;height:1080px} #ov{position:absolute;inset:0}.title{position:absolute;left:84px;top:110px;font-size:118px;line-height:1.05;letter-spacing:-5px;font-weight:650;color:#0a211d}</style></head> <body><div id="lab" data-composition-id="lab" data-width="1920" data-height="1080" data-duration="3.6"> <canvas id="c" width="1920" height="1080" data-layout-allow-overflow></canvas> <div class="title">Your title<br>in context.</div><div id="ov"></div></div> <script type="module"> import { createScene } from './assets/my-scene.js'; import { mountLabels, updateLabels, OVERLAY_CSS } from './projected-overlays.js'; const DURATION = 3.6; document.head.insertAdjacentHTML('beforeend', `<style>${OVERLAY_CSS}</style>`); window.__hf = window.__hf || {}; window.__hf.buildReady = window.__hf.buildReady || {}; window.__hf.buildReady.lab = createScene(document.getElementById('c')).then(scene => { const labels = scene.anchors ? mountLabels(document.getElementById('ov'), [['a','01','Layer one'],['b','02','Layer two']]) : null; const draw = t => { const a = scene.anchors?.(t); scene.render(t); if (labels && a) updateLabels(labels, t, a, 1.2); }; window.__timelines = window.__timelines || {}; const st = { t: 0 }; window.__timelines.lab = gsap.timeline({ paused: true }).to(st, { t: DURATION, duration: DURATION, ease: 'none', onUpdate: () => draw(st.t) }); draw(0); window.addEventListener('hf-seek', e => draw(e.detail.time)); }); </script></body></html>business-motion-film/templates/projected-overlays.js4.8 KBff0cf56d2427upstream ↗
// HTML/SVG overlays driven by projected Three.js anchors. Pure functions of local time (seek-safe). // Labels: an evenly spaced column of callouts with leader lines to moving 3D points. // Pins: photo cards that pop out of a 3D point, dock into a rail, then fly to exact target rectangles in the next scene. // anchors = [{id, x, y}] in screen pixels (from vector.project(camera)). const clamp=(x,a=0,b=1)=>Math.min(b,Math.max(a,x)); const ease=x=>{x=clamp(x);return x*x*(3-2*x)}; const out=x=>1-Math.pow(1-clamp(x),3); const lerp=(a,b,t)=>a+(b-a)*t; const SVG='http://www.w3.org/2000/svg'; function line(svg){const p=document.createElementNS(SVG,'path'),c=document.createElementNS(SVG,'circle'),r=document.createElementNS(SVG,'circle');r.setAttribute('r',16);r.setAttribute('class','pin-ring');c.setAttribute('r',7);svg.append(p,r,c);return{p,c,r};} // items: [[id, number, label], ...] export function mountLabels(root,items){ const svg=document.createElementNS(SVG,'svg');svg.setAttribute('class','ov-lines');root.appendChild(svg); return items.map(([id,n,name])=>{const d=document.createElement('div');d.className='ov-label';d.innerHTML=`<b>${n}</b><span>${name}</span>`;root.appendChild(d);return{id,d,...line(svg)};}); } export function updateLabels(els,t,anchors,start=1.25){ const rows=anchors.map((a,i)=>({a,i,ly:a.y})).sort((p,q)=>p.ly-q.ly);const y0=rows[0].ly,y1=Math.max(rows.at(-1).ly,y0+3*84),gap=(y1-y0)/3;rows.forEach((r,k)=>r.ly=y0+k*gap); rows.forEach(({a,i,ly})=>{const e=els.find(x=>x.id===a.id);const on=out((t-start-i*.11)/.32);const lx=Math.max(a.x+90,1500); e.d.style.transform=`translate(${lx+14}px,${ly-18}px)`;e.d.style.opacity=on;e.d.style.clipPath=`inset(0 ${100-100*on}% 0 0)`; e.p.setAttribute('d',`M${a.x} ${a.y}L${a.x+(lx-a.x)*.45} ${ly}L${lx} ${ly}`);e.p.style.strokeDasharray=700;e.p.style.strokeDashoffset=700*(1-on); for(const c of [e.c,e.r]){c.setAttribute('cx',a.x);c.setAttribute('cy',a.y);}e.c.style.opacity=on;e.r.style.opacity=on*(1-ease((t-start-i*.11-.2)/.5));e.r.setAttribute('r',8+22*ease((t-start-i*.11)/.7));}); } // photos: [{id, n, name, src, at, target:[x,y,w,h], offset:[dx,dy]}] // at = local time the card pops out; target = exact rectangle in the next scene; offset = card position relative to its pin. export function mountPins(root,photos){ const svg=document.createElementNS(SVG,'svg');svg.setAttribute('class','ov-lines');root.appendChild(svg); return photos.map(({id,n,name,src,at,target,offset})=>{const d=document.createElement('div');d.className='pin-card';d.innerHTML=`<img src="${src}"><span><b>${n}</b>${name}</span>`;root.appendChild(d);return{id,at,target,offset,d,...line(svg)};}); } export function updatePins(els,t,anchors,{handoff=3.45,D=4.1}={}){ const W=440,H=248,RW=384,RH=216; anchors.forEach((a,i)=>{const e=els.find(x=>x.id===a.id);const on=out((t-e.at)/.38); const [ox,oy]=e.offset;let px=Math.min(Math.max(a.x+ox,70),1420-W),py=Math.min(Math.max(a.y+oy,70),1080-H-70); const dock=out((t-e.at-.72)/.42),leave=out((t-handoff-i*.06)/(D-handoff-.12)); const rail=[1466,150+i*262,RW,RH];const [tx,ty,tw,th]=e.target; let x=lerp(px,rail[0],dock),y=lerp(py,rail[1],dock),w=lerp(W,RW,dock),h=lerp(H,RH,dock); x=lerp(x,tx,leave);y=lerp(y,ty,leave);w=lerp(w,tw,leave);h=lerp(h,th,leave); e.d.style.transform=`translate(${x}px,${y}px) scale(${lerp(.72,1,on)})`;e.d.style.width=w+'px';e.d.style.height=h+'px';e.d.style.opacity=out((t-e.at)/.12); e.d.querySelector('span').style.opacity=1-leave;e.d.style.boxShadow=`0 22px 50px rgba(0,0,0,${(.4*(1-leave)).toFixed(3)})`; const lineOn=on*(1-dock);const cx=x+w/2,cy=y+h; e.p.setAttribute('d',`M${a.x} ${a.y}L${cx} ${cy}`);e.p.style.opacity=lineOn;for(const c of [e.c,e.r]){c.setAttribute('cx',a.x);c.setAttribute('cy',a.y);}e.c.style.opacity=lineOn; e.r.style.opacity=lineOn*(1-ease((t-e.at-.15)/.55));e.r.setAttribute('r',8+26*ease((t-e.at)/.7));}); } export const OVERLAY_CSS=`.ov-lines{position:absolute;inset:0;width:1920px;height:1080px;overflow:visible;z-index:4;pointer-events:none}.ov-lines path{fill:none;stroke:currentColor;stroke-width:2}.ov-lines circle{fill:#d9ef71;stroke:#0a211d;stroke-width:3}.ov-lines .pin-ring{fill:none;stroke:#d9ef71;stroke-width:2} .ov-label{position:absolute;left:0;top:0;z-index:5;display:flex;align-items:center;gap:12px;color:#0a211d;font-size:26px;font-weight:650;letter-spacing:-.3px;white-space:nowrap}.ov-label b{font-size:17px;letter-spacing:2px;color:#f4f0e5;background:#0a211d;padding:6px 9px;font-weight:600} .pin-card{position:absolute;left:0;top:0;z-index:6;transform-origin:50% 100%;overflow:hidden;background:#0a211d;box-shadow:0 22px 50px #0006}.pin-card img{width:100%;height:100%;object-fit:cover;display:block}.pin-card span{position:absolute;left:12px;top:12px;background:#f4f0e5;color:#0a211d;font-size:19px;font-weight:650;padding:6px 10px;display:flex;gap:10px;letter-spacing:.3px}.pin-card span b{color:#51703a}`;README.mdupstream ↗
Contents are not inlined in the catalog. Open the upstream file.
Plain-text copy, no login: /skill-bundles/echris6-motion-video-kit-business-motion-film.md