Video Breakdown: Shots, Prompts and Remake Plans
Analyze each shot to identify how it was made, edited and scored, and which assets it needs. rea-video combines a CLI with an agent workflow to produce a visual report, prompts and a remake plan.
Works with Claude Code · Codex
Requirements
- Python 3.10+
- ffmpeg / ffprobe
- Git
Overview
What you get
- An HTML/Markdown report, shot contact sheets, a storyboard, audio analysis and subtitle/overlay timing.
- Per-shot evidence, production-method judgments and remake routes, refined through an image-capable agent’s review or an optional video model.
- Separate plans for filming, AI generation, code animation and supplied footage; optional capsule-cinema integration and timeline assembly.
Tutorial
Turn a reference video into a production plan
You find a video you like but cannot tell which shots need filming, which can be generated, or where captions and sound effects belong. rea-video detects shots, extracts frames and audio information, then uses an image-capable agent’s review to produce an evidence-based report and remake plan.
Its main deliverables are analysis and planning materials. Production comes next: you supply real footage, coding tools build motion graphics, and capsule-cinema can optionally generate AI shots.
Choose between rea-video, Cook and Stageshow
- Cook creates and refines a piece against reference images or videos, focusing on making, comparing and improving it.
- Stageshow rebuilds product interfaces and cursor interactions in React for product demonstrations.
- rea-video first studies shots, camera work, editing, sound and assets, then proposes how to make each part.
The project borrows the evidence-first idea from the software reverse-engineering tool REA. It is independent and does not require REA to be installed.
Observe shot structure and element swaps
This page includes two official videos:
| Video | What to observe |
|---|---|
| Four-way outfit comparison, about 12 seconds | Left to right: original, OiiOii remake, author remake, Hanfu/study variation. Compare changed people, clothes and settings with the retained action and transition rhythm. |
| Mixed-media clip, about 19 seconds | AI characters alternate with code-built UI and real screen recording. One video can use several production methods. |
The comparison was made from the analysis using capsule-cinema and generation models. It is not an output rendered by rea-video alone.
Open the official mixed-media report. Compare the shot table, generation units and asset list with the clip. The author records ten shots and divides the underlying visuals between code/template rendering and AI video.
The repository sample is recompressed to 480×854; the report describes the original 1080×1920 analysis input. Use the sample to learn the report and review workflow; its analysis values may differ.
Install the skill and its CLI
Prepare Python 3.10 or later, Git, ffmpeg and ffprobe. This tutorial uses Claude Code or Codex with image inspection for review. No separate video-model key is required for the first exercise.
- Run in a terminal:
# Clone the repository, then install the CLI and skill
git clone https://github.com/JuneYaooo/rea-video
cd rea-video
bash scripts/install_skill.sh
rea-video --help
The script installs Python dependencies and links the skill into existing ~/.claude, ~/.codex or ~/.agents directories. If system Python blocks installation, it tries pipx or a repository-local .venv. Keep the repository directory because the installation uses its path.
- Confirm that
rea-video --helpdisplays its commands. If the script reports that~/.local/binis missing from PATH, add it as directed and reopen the terminal. - Reopen your coding agent and ask it to read the rea-video skill. Supply the working CLI path if the agent’s environment differs from your terminal.
Audio separation downloads a roughly 67 MB model on first use. An optional external VLM is a multimodal model service that can inspect video and audio. Follow the setup reference when you need one.
Run a first breakdown and inspect the report
- Start with the bundled mixed-media sample. From the repository directory, run:
# Produce an English report without an external VLM
rea-video decompose docs/cases/videos/dual_role.mp4 -o out/dual-role --lang en
rea-video summary out/dual-role/breakdown.json
- Open
out/dual-role/report.htmlin a browser. Inspect the shot order instoryboard.jpgand the report’s production methods, editing layers, sound and asset list.shots/contains first/middle/last-frame contact sheets;stems/contains separated audio. - Give the output directory to an image-capable coding agent and send this practice request, written for this page:
Use the rea-video skill to review [absolute path to out/dual-role] and update the report in English.
Inspect the storyboard, then contact sheets for low-confidence shots. Judge production methods independently.
Run review to create tasks; fill review_answers.json with evidence and judgments.
Leave unseen shots unanswered and unconfirmed audio sources unresolved.
Merge answers with apply-review, read the new summary, and explain major changes and remake routes.
Analyze only for now; do not generate a video.
The agent uses these official commands:
rea-video review out/dual-role
# After the agent inspects images from review_tasks.json and fills review_answers.json
rea-video apply-review out/dual-role --by agent
rea-video summary out/dual-role/breakdown.json
Check the updated report for review evidence and reasons behind AI, live-action or code classifications. A cartoon 3D appearance alone is insufficient: the agent should examine deformation, flicker and geometric consistency. Use the updated summary rather than the earlier proportions.
Without an external VLM, an image-capable agent can review still frames. It cannot establish voice, music or effect sources from audio it has not heard. Listen to the separated tracks or connect an audio/video-capable model when those questions matter.
Plan a replacement for one shot
Before remaking a whole video, choose one shot and specify what to retain and replace. This practice request was written for this page:
Read the reviewed report and plan a product replacement for shot [number] only.
Product reference image: [absolute image path].
Keep the camera position, movement direction, duration and transition, but use my product's appearance.
Separate footage to film, visuals to generate and captions to add in post-production.
List assets, a prompt and a preview plan for this shot. Do not start paid generation.
Check that the product has a matching reference image and that movement can connect to adjacent shots. Then try a local revision: “Change only the product color. Keep the person, setting and shot duration.” This keeps one edit tied to one clear element.
Make captions, styled text and stickers as separate editing layers rather than baking them into generated footage. When the script changes, retime captions and sound effects to match the new dialogue.
Continue to generation or assembly when ready
AI-shot generation requires a separate capsule-cinema installation, its Python environment and model-service configuration. Ask the agent to prepare jobs using the official workflow, then run rea-video remake-run [path to jobs.json] --dry-run to inspect tasks and costs before approving execution.
The report can group several shots into one generation unit. Current automatic execution still calls the model once per shot; planned grouping does not establish actual call counts.
Once you have filmed or generated replacement shots, rea-video assemble can fit them back onto the original timeline. Missing replacements retain original footage as placeholders, and audio defaults to the original track. Use the assembly log to fill those gaps before delivering your own complete work. See the official remake reference for parameters and asset mapping.
Apply the workflow to your own reference
After completing the sample, replace the paths and goals in this template:
Analyze with rea-video: [absolute local video path].
Report language: English.
Purpose: [study camera work / adapt a product ad / rebuild code animation].
Focus: [shots, camera movement, captions, sound, etc.].
Keep: [pacing, composition or a specific action].
Replace: [people, product, setting or wording].
Available assets: [paths to images, recordings, footage and audio].
Produce and review the report first. Then separate filming, generation and coding tasks, and wait for me to select the first shot.
Begin with a short clip or one section. Confirm its production route and assets before expanding, so you can decide which effects to retain and which need your own filming or implementation.