RoboColosseum Field Evaluation Launch

A documentary launch explainer opens on a real task, turns it into a measurement problem, then reveals a product through field proof, a rapid montage, a working process, comparative data and a concrete offer. Its engine is spoken argument with visible evidence at each claim.

Why it works

• The opening task is readable at frame 0 before the product is named.
• Shots 2–10 compress many kinds of proof into the 27.60–31.84 s passage.
• The final shot combines a human presenter, large offer type and an actionable close.

Tools you'll need: a camera shoot · Remotion for motion graphics · ElevenLabs or a recorded voice · Remotion or any video editor · Claude alone: no

Hook formula (inferred): observable task → evaluation question: “How can we tell whether {{PRODUCT}} works?”

Why it holds (inferred): The audience can understand the action without context; the visual scope grows while the voice sharpens the question.

Retention map (inferred): re-hooks at beats 2, 4 and 8; payoff and action in beat 9; hard stop, no fade to black.

Mix (inferred): One long composite opening, a nine-shot burst, then presenter/field/data composites. Rhythm: Speech carries the long sections; the burst interrupts them. The last shot is a long graphic build, not one static card.

Rules this editor follows (inferred):

• Begin and return to real footage when introducing claims.
• Keep one compressed montage at beat 4; leave the explanatory passages room for speech.
• Use the keyword overlays as phrase punctuation, never full-time subtitles.

Look (inferred): Practical office and field light, mostly neutral footage. Dark information surfaces and one saturated accent isolate the claims; map those roles to {{BRAND_BG}} and {{BRAND_ACCENTS}}. Graphics: thin frames, small labels, simple rectangular highlight strips. Transitions: cuts where plate 5 lists cuts; continuous motion at the two other boundaries. Branding: end on {{LOGO}} and {{CTA_URL}} as text.

Grade and shoot card (inferred)

Shoot 4K so a 1920×1080 crop survives the shot 11 push. In the 2D comp, use wide frames for the opening task and environment, medium presenter crops and close device details; keep the task readable. At 1080p, keep the measured full-frame focus box (0 · 0 · 100 · 100%) sharp with 0 px blur outside it in shots 1–12; at frame 1167 (46.68 s), ramp the shot 12 footage blur over 8 frames until edge sharpness reaches ×0.3, while keeping captions sharp. Use a soft highlight glow of 12 px radius at 15% opacity and a restrained edge vignette of 20% opacity over the outer 12% of the frame. Neutral grade: modest contrast, saturation about 0.9, slightly raised shadows, restrained highlights. Record clean presenter lav/boom, 30 s room tone per location, and task sound. Capture one clear task, two presenter setups, at least nine distinct proof clips, one process demonstration, six trial views and a clean closing background. Use {{BRAND_BG}} for data cards and {{BRAND_ACCENTS}} for thin borders/highlights; no 3D is needed.

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

Measured: 225 words and 73% speech coverage per plate 8; v1 notes estimate 208 wpm while speaking and a 4.56 s speech gap beginning near 26.9 s.

Voice direction (inferred): A capable guide explains one specific uncertainty to peers. Use a curious opening question, precise nouns for the process, and a firm, clean offer. Leave beat 4 to picture and music. Different presenters can share the script, but do not imitate the people in the reference.

Music (inferred): Restrained tonal electronic bed under speech, slightly more present during beat 4, thinning beneath the final words. Effects: very sparse interface ticks for verified card changes; avoid the strong cues marked “voice?” in the measured sheet. Space: keep the proof burst free of narration, with no full-mix silence.

██▇█▇▇▇██▇ ▇█████████ ██████▇▇█▇ ▇▇████████ ████▇▇▇███ ██▇██▇████ ███▇███▇█▇ ███▇████▇▇ ██▇▆▆▆▄▁

Title (inferred): “Introducing {{PRODUCT}}: evidence from real use.” Cover: a readable physical task, subject to one side and room for a short headline. Post copy: name the launch, cite only verified scope and the specific offer, then add {{CTA_URL}}. Audience signals: the source post had 115,773 views at capture; no replay or comment-level evidence is supplied.

Narration-led (inferred): Roughly 225 words at about 208 spoken wpm, with the beat 4 gap. Cut at phrase endings; captions highlight only the claim word or number.

[beat 1 · shot 1 · ~18 words] Describe a real task, then ask how {{PRODUCT}} can be evaluated.
[beat 2 · shot 1 · ~35 words] Name the inconsistent outcome; show only verified examples.
[beat 3 · shot 1 · ~24 words] Introduce {{PRODUCT}} and the real-world scope.
[beat 4 · shots 2–10 · 0 words] Let the proof clips carry the claim.
[beat 5 · shot 11 · ~22 words] Explain why settings differ.
[beat 6 · shot 12 · ~27 words] Describe input, real process and observable output.
[beat 7 · shots 13–16 · ~29 words] Explain the method and its limits.
[beat 8 · shot 17 · ~24 words] Compare only defensible results.
[beat 9 · shot 18 · ~45 words] State availability, terms and one action; end on {{LOGO}} and {{CTA_URL}}.

Two routes (inferred)

Minimum: Remotion end to end for edit, graphics, captions and Claude-synthesised music. User supplies real footage, verified figures and a recorded or TTS voice; Claude alone cannot capture them. This reaches the timing and typography while simplifying field coverage to available clips and giving up the reference's breadth of locations.

Full: Shoot 4K footage and record presenter/room audio → clean and select clips → record voice or produce a licensed ElevenLabs voice WAV → edit and graphics in Remotion → mix original bed → H.264 master. Handoffs: camera originals to edit as 4K clips at 25 fps; shot selection and camera keyframes as JSON; graphic overlays as 1920×1080 RGBA PNG sequences at 25 fps when compositing externally; voice/music/room tone as 48 kHz WAV, with a separate click-track WAV aligned to cue frames; final 1920×1080 at 25 fps.

Asset honesty: real captures of {{PRODUCT}} or its UI rebuilt in code; where none exists, label a type tile plainly. Never generated product UI, invented screenshots, fake metrics, fake reviews or fake outputs. Any comparison bar in beat 8 needs a traceable measurement.

Script prompt (inferred)

Write roughly 225 spoken words for a narration-led launch film for {{PRODUCT}}. Follow beats 1–9 in plate 4 and the phrase space in plate 10. Start with a real task and a testable question; leave beat 4 unspoken. Supply verified footage and figures before writing any numeric claim. End with one concrete action and {{CTA_URL}}. Do not borrow the reference's wording, brand or numbers.

Caption and music prompts (inferred): Animate only the few claim words in plate 10 using {{BRAND_FONT}}, light/dark roles from {{BRAND_BG}}, and highlights from {{BRAND_ACCENTS}}. Music prompt is in plate 8.

How it's built (inferred)

Signature move: the proof burst and hand-off. Keep shot 1's spoken promise until frame 689, then hard cut at the plate 5 boundaries through shots 2–7. These shots are short by design; use footage at native speed, no interpolated dissolve. Shots 8 and 9 are continuous at frames 748 and 772: maintain one underlying clip or matched camera angle and use an overlaid moving crop/window for the transition, 10 f each. Plate 5's shot 8 region appears to move far right and enlarge; confidence is 0.30, so use it as a visual hint, not a literal camera command. The sheets show the same room/robot lineup through frame 748 and a continuous drift toward the table before shot 9. Ease the overlay cubic-bezier(.2,.8,.2,1), with 180° shutter motion blur at 25 fps; avoid doubling a subject at the mask seam. The next hard cut is frame 784.

Other moves. In shot 1, the opening task collapses into a collage around 1.50–3.00 s; this is an internal compositing build, not a measured cut. Build a 2×2 proof grid, then a dense thumbnail wall, with 6–10 f ease-out scales and masked edges. In shot 11, hold the base frame to 35.28 s, then apply the measured ×1.29 image push and +5.4% x travel over 10 f to 35.68 s; cubic-bezier(.42,0,.58,1), 180° shutter blur. This is a composited image reframe, not a dolly. In shot 12, follow the element's measured motion segments and pull focus beginning frame 1167 over 8 f to edge sharpness ×0.3; use a blur ramp on the affected footage, keeping necessary captions readable. In shot 13, use the plate 5 1.9% x and −1.0% y element drift and 2.6° roll over its shot, with the direction change at 49.88 s. In shot 17, animate only the data region according to plate 5's segments; score bars must reflect verified values. Shot 18 holds a fixed camera while headlines, counter and CTA replace each other; do not fade the picture to black. Animate on ones, matching the measured 23.3 unique moving frames per second. Keep motion blur on transformed overlays, not on readable type.

Sheet reading (inferred): The opening tile montage around 1.50–3.00 s and the end-card switch near 1:24 are changes within shots 1 and 18 that the transition list does not enumerate. The frames suggest a new graphic state; the measurement says those are within the same measured shot, so keep their durations tied to the visual build rather than inventing cuts.

Build order

1. Assemble verified evidence and record narration for beats 1–9, leaving beat 4 open.
2. Cut footage to shots 1–18 and the plate 4 boundaries; make the shot 8–9 hand-off first.
3. Build shot 11's reframe, shot 12's focus pull, then beats 2, 6 and 8 graphics.
4. Add beat 9 typography and the end card; mix the original bed around stressed words.
5. Export and run plate 14 checks.

• Every hard cut is at a plate 4 frame within ±1 f; frames 748 and 772 remain continuous.
• The shot 11 push reaches ×1.29 over its measured 10 f segment; shot 12 focus softens over 8 f from frame 1167.
• Every essential word is ≥38 px cap at 1920×1080, or the aspect-specific minimum in plate 15.
• The nine proof clips in beat 4 are distinct, with actual task or device evidence.
• Cues follow phrase/picture events, not a BPM grid; no effects cover stressed words.
• X/LinkedIn master is about −14 LUFS integrated, true peak ≤−1 dBTP.
• Every metric, progress display and offer matches supplied evidence; final URL is readable.
• No reference colours, font, logo, slogan, people or music remain.

Brand slots

What this film avoids (inferred): continuous subtitles, decorative transitions, unverified performance claims, stock proof, a detached feature list, QR codes and app-store badges.

Brand-distance warning (inferred): A robot field-test story with orange highlight strips, the same giant offer word and a circular target logo will read too close to the source. Change the primary visual motif to your own verified evidence and brand layout.

Adapt the aspect (inferred):

• 9:16: Re-lay out presenters above/below the proof and stack metric cards; centre crop retains only 10 of 91 measured boxes. Keep readable type ≥22 px cap at 1080 px width, and place the action/URL clear of top profile and bottom control zones.
• 1:1: Re-lay out the wide collage and comparison bars into stacked panels; centre crop retains 25 of 91 boxes. Keep words ≥22 px cap at 1080 px width and avoid the lower platform caption area.

Adapt the length (inferred):

• 15 s: Keep beat 1 as a compact question, a shortened but distinct beat 4, one process proof from beat 6 and the beat 9 action; drop optional setting and extra field inserts.
• 30 s: Preserve beats 1, 3, 4, 6, 8 and 9; shorten narration and keep only the strongest verified example from the optional proof shots.
• 60 s: Add breathing room to the explanation and verified process in beats 2, 5–8; introduce new evidence rather than stretching the burst or end card.

The creator's signature (don't copy): the source's presenters, robot fleet, proprietary evaluation data, name, palette and mark. The transferable format: physical proof → measurement question → spoken reveal → compressed proof burst → demonstrated process → defensible comparison → clear action.

Source: https://x.com/RoboColosseum/status/2117982605274710483 · timeline v2 measured 2026-11-06 · (measured) = read off the film · (inferred) = construction reading