Ever watched an ad and wondered "how did they make that?" Not the story — the production. Which shots were filmed, which were AI-generated, why the cuts land where they do. That question used to require guesswork. It doesn't anymore.
This is a practical guide to breaking a video down into its production decisions — and rebuilding that knowledge into something you can make yourself.
1. Separate the picture from everything on top of it
Most "video breakdowns" stop at the story. A production breakdown starts by splitting the video into layers:
- The base picture of each shot: filmed with a camera, generated by an AI model, rendered in 3D, drawn by code, or a screen recording.
- The packaging: captions, titles, stickers, picture-in-picture — everything composited in editing.
- The sound: voice, music, and effects, which usually follow their own timeline.
The biggest mistake beginners make is treating a shot as one thing. A single two-second shot can be an AI-generated background, a real product photo composited on top, and a caption added in post — three production routes in one moment, each needing different tools.
2. Judge each shot's source like a detective, not a fan
How do you tell filmed from generated? Watch for:
- Physics tells: generated motion often smears details, softens hands, or drifts lighting between frames.
- Cut density: AI clips tend to be single continuous shots (generation limits), while filmed sequences cut more freely.
- Consistency: does the product/logo/text stay identical across shots? Generation often mutates it; filming can't.
Certainty matters. Good analysis states confidence: "likely AI (70%) — hands blur during the turn" beats a confident wrong guess.
3. Read the edit like a rhythm section
Shot lengths tell you the genre's grammar: product ads cut every 1.5–3 seconds; tutorial content holds 5–8; drama leans longer. Note where cuts land relative to sound — a sound effect hitting exactly on a product reveal is an editing decision, not luck.
4. Turn observations into a remake plan
For each shot, assign a production route:
- Film it → note framing, lighting, duration
- Generate it → draft a prompt covering subject, motion, camera, style
- Build it in code (UI demos, text animations) → note the layout and motion
- Assemble it (captions, stickers, music) → list what lands in post, at what timestamp
Keep prompts per generation unit, not per video — most AI video models handle one shot or one short sequence per call.
5. Doing it programmatically
This workflow is exactly what rea-video automates. It's a free, open-source (Apache-2.0) analyzer that runs locally and works as a skill inside AI coding agents like Claude Code or Codex:
# Install per the repo docs, then decompose a clip
rea-video --env-file .env decompose reference.mp4 -o out/ref --vlm --lang en
rea-video summary out/ref/breakdown.json
It labels every shot live-action vs AI-generated vs code-rendered with the evidence frames behind each call, splits captions and audio layers, and generates the remake plan (which shots to film, which to generate, with draft prompts). Your reference videos never leave your machine.
6. Test one shot before committing to the rest
Rebuild your most important single shot first. If the remake reads wrong — timing off, motion unconvincing — adjust before scaling the workflow to all 30 shots.
What you can't recover
Honesty section: compression destroys most forensic traces. You will rarely recover an exact original prompt or an editable project file. What you can recover is the decision structure — and that's the part worth learning anyway.
The workflow in one line: layer it, source it, time it, route it, test one shot.
Top comments (0)