Pre-Production: Invest Once, Use for Every Video
Pre-Production Checklist (Check First, Then Prepare What's Missing)
| Item | Completion Criteria | Reference |
|---|---|---|
| ① Digital Human | One avatar + look; green screen sample passes; clean edges after keying; lip-sync timing calibrated; resolution tested | references/digital-human.md, references/voice-and-host.md Part 2 |
| ② Voice | Cloned voice; pronunciation rules (abbreviations, erhua, near-homophones); pause cap parameters | references/voice-and-host.md Part 1 |
| ③ Motion Effect Patterns | At least one pass through the pattern layer: know two easing levels, staggered timing, click feedback, transitions, restrained post-processing | references/principles.md; UI cards into real shots, real-scene transitions see references/vlog-ui-fusion-and-transitions.md |
| ④ Component Library | Personal Remotion component library: titles, cards, charts, icons, windows, transitions; each with parameters and calibration records | references/method-dissect.md, references/method-replicate.md |
| ⑤ Fixed Framework | Caption capsule style, backgrounds (light/dark each), avatar frame, cover template | references/frame-kit.md |
| ⑥ Fallback Path | Know what to do when Remotion cannot handle something (3D, glass, real-shot compositing) | references/fallback-ladder.md |
Once ①–⑤ are ready (for pure motion-effect videos, only ②③④⑤) → hand off to ../talking-video-oneshot/SKILL.md.
Building Motion Effect Blocks: The Loop for Learning Effects
- Grab Footage: User provides a beautiful motion effect reference (screen recording/video); check specs with
ffprobe; for variable frame rates, create a constant frame rate proxy first; keep original read-only. - Deconstruct: Scan keyframes to locate events → dense frame extraction at 12–30 fps per event → measure displacement / scale / blur / brightness → fit easing curves → write technique cards, annotating each value as "measured / approximate" with confidence level.
- Catalog: First check if existing blocks can be parameterized or combined; only create new ones if none fit. New patterns go into the pattern library: rule → value → source → confidence.
- Replicate: Create a calibration composition; render with card parameters once, measure once; pass if deviation is within ±15%. Only independent pixel inspection proves the original values are correct.
- Side-by-Side Calibration: Show original on left, our version on right; write the gap and "how similar (X%)"; submit to user for feedback and adjust accordingly.
- Record Even Disproven Hypotheses: Conclusions disproven by measurements go into the pattern library under "disproven conclusions" to avoid repeating mistakes.
Replication is for internal learning only — never goes into final output, never copies others' work.
Specifications
- Components written at 1920×1080 pixels; determinism: no random numbers / time, use seeded generators for randomness, derive time from frames.
- One camera per shot; no explanatory small text; no fabricated numbers (if unclear, write "unclear").
- Render one at a time on low-memory machines.
- Communicate with user in Chinese; use multiple-choice questions for decisions that need approval.