Repository navigation
00 characters
github-actions[bot] edited this page Oct 22, 2026
·
1 revision
Every scene is rendered separately, then clipped together. These sheets are the contract that keeps the cast identical across renders. If a render disagrees with this document, the render is wrong.
One character for the whole video. He is the visitor: a home-gym trainee who types his own workouts. No name, no dialogue, no mascot energy.
- Build: simple flat figure, ~5.5 heads tall, rounded limbs, no fingers, no facial detail beyond two dot eyes and a short line mouth. Silhouette must read at 10% scale.
- Outfit (never changes): dark charcoal tee, app-accent shorts, plain trainers. No logos, no headband, no watch (a watch would fight Scene 5's "phone stays down").
- Skin/hair: mid-tone skin, short dark hair block. Keep abstract; he is an avatar of the user, not a portrait of anyone.
-
Standard poses:
- P1 — seated at laptop, three-quarter view, typing. (Scenes 1–2)
- P2 — standing, phone in one hand at waist height. (Scene 5 open)
- P3a / P3b — burpee pair: standing tall / prone on floor. Two frames only, looped; never in-between frames. (Scene 4)
- P4 — close hand holding the remote, thumb on the split button. (Scene 5)
- P5 — standing facing the TV, back three-quarter to camera, mid-set stance. (Scene 5)
- Don'ts: no smiling/frowning arcs (mouth stays a line; emotion comes from posture — the Scene 1 slump is a 4° shoulder drop, nothing more). No extra characters ever. No photoreal rendering of the athlete in any scene.
- A simple cartoon hand holding a thick marker; enters from the nearest board edge, writes/underlines, leaves. It annotates; it never operates the app (that's the Cursor's job) and never touches the Athlete.
- Marker ink: one warm red-orange (distinct from the app accent — see UI guide) so hand-drawn emphasis can never be mistaken for UI. Stroke width and slight wobble identical in every scene; max 3 annotations per scene.
- Standard OS arrow, rendered at 150% with a soft shadow. Every click, in every scene, is identical: arrow settles → one accent-colored ripple ring expands and fades → the app's click state. Same ripple, same 300ms timing, same soft click sound.
- The cursor never types (text types itself, keystroke sound only) and never appears in gym scenes (5) — there, hands and remotes act.
- PROP-1 Laptop: thin silver slab, screen = UI screens; wod.wiki screens only, generic apps only in Scene 1.
- PROP-2 Phone: black slab, minimal bezel; states: editor, timer, "casting" dim.
- PROP-3 TV: wall-mounted rectangle, thin black frame; shows UI-2 at scale.
- PROP-4 Remote: small dark wand, few buttons. One button — the split button — carries a permanent app-accent ring. Same button, same ring, every frame, every render. This prop is effectively a character; treat changes as casting changes.
- PROP-5 Bluetooth keyboard: compact, no numpad, space-gray; appears only in Scene 5's inset beat, same angle both beats.
- PROP-6 Kettlebell silhouette: one, in the gym wide shot, as set dressing only.
- Sergei narrates all 7 scenes in his own voice. No AI voice, no hired VO.
- Register: the founder explaining his own product — plain-spoken, coach-engineer, conversational pace (~150 wpm), zero hype-adjectives, no "revolutionary."
- The consistency risk moves from voice choice to recording setup: same microphone,
same room, same distance and gain for every scene; record all seven scenes in one
session where possible, and log the setup here before the first take:
MIC/SETUP: TBD — fill in at recording. - If a scene is ever re-recorded later, match the logged setup and A/B the first line against the original take before recording the rest.
- Narration scripts in the scene files are written to one sentence rhythm on purpose — preserve it when editing, and re-time the scene if the word count changes.