A Robot Failure Becomes a Product Proof

A physical failure opens a founder led explanation of its field cost. The film then shows a real input, a controlled test, a fix, and fresh validation before turning proof into a buyer outcome.

Why it works

• The failure is legible before the first cut at frame 26.
• The product proof starts with the simulation handoff at frames 272–305, then moves to real interface evidence in shots 13–16.
• Oversized evidence cards in shots 20 and 23 make the later results readable even with sound off.

Tools you'll need: a camera shoot · screen recordings of the product · Blender for 3D · ElevenLabs or a recorded voice · Remotion for motion graphics · Claude alone: partly

Hook formula (inferred): visible failure + direct question: See {{PRODUCT}} fail at {{SITE}}?
Why it holds (inferred): The audience sees the action before it knows the explanation. The camera change at frame 26 makes the fall spatially clear.

Retention map (inferred): re-hooks at shots 9, 20 and 23 · validation at shot 28 · call to action at shot 31 · hard stop with no end fade.

Mix (inferred): Physical simulation and footage anchor the argument; live presenters explain; UI and flat cards carry proof.
Rhythm: The measured median is 2.21 s. Shot 10 is the long explanatory hold; shots 20–28 compress into short result cards and one continuous card.

Rules this editor follows (inferred):

• Cut on clause boundaries. The beat grid is secondary to spoken emphasis.
• Keep a moving subject inside a stable composition before revealing the outcome.
• Alternate real presenter, simulated action and captured interface so each claim gets a distinct kind of evidence.

• Look (inferred): The reference moves between warm dusty simulation, dim warm interviews, nearly black UI, pale editorial cards, and saturated proof cards. Map those roles to {{BRAND_BG}} and {{BRAND_ACCENTS}}.
• Graphics and overlays (inferred): Flat pixel icons, dot grids and oversized numerals sit in clean fields; important words appear in short phrase groups.
• Transitions (inferred): Mostly cuts; preserve the measured whip, zoom and white fade only where they clarify a handoff.
• Branding and end card (inferred): Shot 31 changes picture at frame 2157 to a plain {{LOGO}} and {{CTA_URL}} text card; it stays visible through the final frame without a fade.

Grade and shoot card (inferred)

Shoot two presenter clips, a real site context clip, and authentic {{PRODUCT}} screen recordings. Presenter: 35–50 mm full-frame lens at eye height, level camera, chest-up; f/2.8–4 for soft corridor depth. Key: 4500 K large soft source 45° off-axis; fill 1.5–2 stops under, practical warm background. Grade: moderate contrast, lifted shadows, restrained saturation, warm midtones; preserve dark garment detail. Capture 4K for the measured reframes and record isolated 48 kHz dialogue and room tone. For simulation, Blender Principled BSDF: painted shell roughness 0.45, metallic 0.15, IOR 1.45, coat 0.2; glass roughness 0.08, transmission 0.85; world strength 0.35, Cycles 64 samples, 35 mm camera at low panel height. The 2D equivalent is layered CSS extrusion of 18 px, blur 6 px outside shot 3's 33 · 71 · 50 · 29 focus box, 24 px warm glow at 0.18 opacity, and a 20% vignette reaching full strength at the frame edge.

:root { --brand-bg: {{BRAND_BG}}; --brand-accent: {{BRAND_ACCENTS}}; }
.proof-card { background: var(--brand-bg); color: white; border: 0; border-radius: 0; box-shadow: none; }
.ui-panel { background: color-mix(in srgb, var(--brand-bg) 88%, white); border: 1px solid rgba(255,255,255,.16); border-radius: 5px; }
.sim-glow { filter: drop-shadow(0 0 24px rgba(255,255,255,.18)); }

      0s         10s        20s        30s        40s        50s        60s        70s        80s        90s
mean  ▃▄▃▂▁▁▁▃▃▃ ▃▂▃▄▂▂▂██▂ ▂▂▂▂▂▂▂▂▂▃ ▄▃▅▃▂▂▁▂▂▁ ▁▁▁▁▄▃▂▅▅▅ ▃▃▄▄▃▄▅▅▃▃ ▂▁▃▄▄▄▄▅█▆ ▃▃▄▄▄▃▃▄▄▃ ▂▃▄▄███▆▂▂ ▁▁▁▁
p95   ██▇▄▁▁▅▇▇█ ▆▄▇▇▆▇▇██▅ ▅▅▄▄▅▄▅▅▅▆ █▇▆██▆▂▅▄▂ ▃▃▃▂▄▅▇███ ▇▅▅▅▅▇██▆▅ ▃▅▆▅▄▄▆██▆ ▅▅▆▆▅▃▃▄▅▅ ▅▆█████▇▄▄ ▁▁▁▁

Measured: The v1 transcript counts 240 words, 156 wpm overall and 204 wpm during speech; plate 8's aligned transcript counts 228 words covering 71%. These methods differ, so budget from the 228 aligned words. The delivery uses frequent short pauses and medium-register, animated pitch.

Voice direction (inferred): Two distinct technical presenter archetypes speak as people explaining their own work. Start with curiosity, move briskly through the field cost, leave air before the large proof numbers, and land the destination plainly. Record each voice cleanly; avoid imitation of the original speakers.

• Music (inferred): A quiet tonal electronic bed supports dense speech. Keep its rhythm perceptible but never steer the cuts to it.
• Sound effects (inferred): A soft physical scrape in beat 1, a short whip at frame 305, restrained interface clicks around the workflow, and low impact accents at the proof cards. Leave cues marked “voice?” as voice, not effects.
• Silence and space (inferred): Thin the bed before beat 7's numbers and let the last voice line clear before the card settles.

█▇▇███▆███ ▇█▇█▇▇▇███ ▇██▇▇█▇▇▆█ ▇▇▇█▇▇▆█▆█ ▇▇▅▇█▆▅█▇▇ ▇██▆▇▇▇▆▇█ ▅▅██▇▇▆▇▇▅ ▄▇█▇▇█▇▇▇▇ █▇▇▆█▇▇▆█▆ █▇▂▁

• Title (inferred): Show {{PRODUCT}} facing {{FAILURE}} before it reaches a customer.
• Cover (inferred): Freeze the near-failure at the edge, one short question in {{BRAND_FONT}}, brand roles from plate 15.
• Post copy (inferred): One field risk, one verifiable test result, then {{CTA_URL}} as plain text.
• Audience signals: The captured post has about 185k views; replay and completion data are unavailable.

The narration leads. Write about 225–240 words for the full cut, with quick spoken runs near 200 wpm and deliberate pauses before proof numbers. Captions follow clauses and stay above the phone minimum.

[beat 1 · shots 1–3 · ~12 words] See {{PRODUCT}} encounter {{FAILURE}}? State the field consequence.
[beat 2 · shots 4–5 · ~21 words] Name the manual response and its cost.
[beat 3 · shots 6–8 · ~17 words] State what {{PRODUCT}} tests and what changed.
[beat 4 · shots 9–11 · ~38 words] Explain the old burden, then its repeat failure.
[beat 5 · shots 12–13 · ~13 words] First-person credibility in one compact sentence.
[beat 6 · shots 14–18 · ~33 words] Speak the real input, code handoff and variable conditions.
[beat 7 · shots 19–27 · ~57 words] State verified scale, failure diagnosis and fix; pause before each number.
[beat 8 · shots 28–30 · ~34 words] Give fresh validation and the buyer outcome.
[beat 9 · shot 31 · ~14 words] Close with a new claim, then {{CTA_URL}}.

Two routes (inferred)

Minimum: Remotion with CSS 3D, code-built UI and cards, and music synthesized in code. Supply real presenter footage, recorded or TTS speech, product captures and verified results. This reaches the evidence rhythm and graphic look while simplifying the robot to layered 3D transforms; it gives up realistic contact, dust and physical light.

Full: Shoot 4K presenters and site footage; record 48 kHz WAV dialogue; capture real {{PRODUCT}} UI; build the failure simulation in Blender at 1920×1080, 23.976 fps; export ProRes or PNG sequence with alpha plus camera keyframes JSON; assemble captions, cards and shots in Remotion at the same size and rate; hand in an original 48 kHz WAV bed with click track and separate effects; export H.264 MP4. The Blender camera JSON and 1920×1080 PNG sequence pass to Remotion; the dialogue WAV and music WAV remain separate until final mix.

Asset honesty: Use real {{PRODUCT}} captures or your own UI rebuilt in code; mark missing proof as missing. Never generate product UI, fake screenshots, fake results, fake reviews or invented metrics. The failure simulation may be illustrative, but the fix and fresh validation must come from a real test.

Script prompt

Write a narration-led launch script for {{PRODUCT}}. Follow plate 4's nine beats and shot groupings. Use a visible {{FAILURE}} at {{SITE}}, the actual buyer cost, a real input, verifiable test conditions, verified result and a plain-text {{CTA_URL}}. Budget about 230 words; use two distinct technical presenter archetypes, short clauses and pauses before large numbers. Output spoken line, caption and evidence source per beat. Never invent metrics or mimic the reference's words.

Caption spec (inferred): Phrase captions follow speech, not a bouncing word effect. Keep required words at least 38 px cap at 1920×1080, with a dark translucent backing over footage and no backing on clean cards. Reserve the large plate 6 levels for editorial claims.

Music prompt: Use plate 8's original-bed prompt and deliver isolated 48 kHz WAV stems.

How it's built (inferred)

Signature failure move: In shots 1–2, keep the panel edge as a fixed spatial anchor and animate the machine as a rigid object with separate track, wheel and shadow layers. Shot 1 runs frames 0–25: the measured image rolls +14.8° while drifting +2.5% x and −6.1% y, accelerating with cubic-bezier(.55,.05,.8,.65). At frame 26 cut to the wider, static camera; finish the rigid-body fall in the object layer. Render animation on ones with 180° shutter, or 8 subframes at 23.976 fps to avoid strobing. Do not attach the shadow to the machine: let it detach from the panel and fall onto the ground.

Evidence handoffs: Shot 4's camera is static while the icon group travels −30.6% x and scales ×0.71; keyframe its measured segments from plate 5. Shot 5 moves from a wide site frame toward a close machine frame; at frame 196 pull focus over 26 f by crossfading two depth layers and doubling edge sharpness, with no new cut. At frame 272 continue the logo-to-simulation picture over 7 f without a visible cut, using a rising panel mask. Shot 7's −86.7% y tilt moves the panel field under the camera; at frame 305 apply the measured 4 f horizontal whip into shot 8. At frame 357 step the light grade from measured luma 0.156 to 0.243 while the split code/simulation view continues; the frame strip suggests a changed simulation image, but the measurement labels the event an unknown brightness step. At frame 456 use the measured 2 f zoom out into shot 10. At frame 780 reveal shot 13 from white over 8 f. At frame 1706 continue the validation graphic over 12 f with a mask/position handoff, not a cut. At frame 2157 make the internal hard cut from the closing claim to the end card; do not fade at the end.

Other motion: Follow plate 5's segmented camera keyframes for shots 8, 13, 15 and 28. In shot 28 the camera stays still while the icon group moves −31.1% x, −16.1% y and scales ×0.63 across its moving portion. The sheet shows a result card becoming a use-case row, so animate independent icons and captions rather than a whole-frame pan. Use cubic-bezier(.2,.7,.2,1) for measured ease-outs and cubic-bezier(.42,0,.58,1) for ease-in-out; where plate 5 says irregular, fit keyframes to the supplied camera JSON instead of imposing one curve. Use directional blur only on the 4 f whip; keep type sharp.

Build order

1. Lock the narration and evidence sources to beats 1–9, with clause markers and verification for every number.
2. Build beat 1's failure in Blender or the Minimum stack, then test the frame 26 spatial handoff.
3. Shoot and record beats 2, 4, 5 and 8; capture real UI for beats 3, 6 and 7.
4. Assemble all shots at the measured frames; apply plate 5's camera and transition data, then add cards and captions.
5. Mix the original bed beneath speech, inspect the cue frames, master, and export.

• Shot boundaries match plate 4 within ±1 frame; frames 272 and 1706 remain continuous.
• Shot 1's +14.8° roll and shot 7's −86.7% vertical travel match plate 5.
• Focus changes for 26 f from frame 196; the frame 357 grade step remains inside shot 8.
• The end-card cut occurs at frame 2157; no fade to black follows.
• All required words meet the 38 px cap phone minimum; decorative UI text is not needed to understand the claim.
• Spoken lines remain intelligible over cues; do not add effects to events marked “voice?”.
• Export 1920×1080, 23.976 fps H.264; X/LinkedIn master near −14 LUFS and true peak ≤−1 dBTP.
• Every displayed result has a verifiable source; no reference identity, footage, copy or music remains.

Brand slots

What this film avoids: no feature-list montage, no stock testimonial, no decorative fake dashboard, no unsupported result, no QR code or app-store badge.

Brand-distance warning: A near-identical solar robot, orange/blue proof cards, dark hoodie presenters and the reference's closing slogan together read as the original. Change the physical failure and site first; then apply your own brand system and copy.

Adapt the aspect (inferred):

• 9:16: Re-lay out: stack presenter above proof, move captions into the middle safe region, keep key text ≥22 px cap at 1080 width, and reserve the top and bottom 15% for platform UI.
• 1:1: Re-lay out the split screens into sequential cards; centre the physical subject and keep text ≥22 px cap at 1080 width, with 10% top/bottom clear for platform chrome.

Adapt the length (inferred):

• 15 s: Keep beat 1's failure, a compressed beat 3, one verified result from beat 7, beat 8's buyer outcome and beat 9; drop optional explanation and secondary examples.
• 30 s: Keep all nine beats but reduce beat 4's manual burden and beat 6's conditions to one example each.
• 60 s: Keep the full workflow and one additional verified condition in beat 7; add evidence rather than lengthening the end-card or hook holds.

The creator's signature (don't copy): The original presenters, robot design, solar setting, UI, wordmark, palette, specific claims and music.
The transferable format (do copy): A visible failure, human explanation, authentic workflow proof, fresh validation and buyer outcome.

Source: https://x.com/alliance/status/2118366494723620968 · measured with dissect 1.1 and timeline v2 on 2026-11-06 · (measured) = observed · (inferred) = construction reading