Communitygithub.com

rea-video

Shows an agent how to plan, build, or cut "rea video" so the concrete output is a logo and a small brand sheet. The procedure moves through rea-video:视频制作方式逆向拆解, 0. 环境, 1. 标准流程. Use it when a new assignment calls for this craft.

rea-video 是什么?

rea-video is a Claude Code agent skill that shows an agent how to plan, build, or cut "rea video" so the concrete output is a logo and a small brand sheet. The procedure moves through rea-video:视频制作方式逆向拆解, 0. 环境, 1. 标准流程. Use it when a new assignment calls for this craft.

兼容平台~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/rafaelleandro111/estrategia-de-videos/tree/HEAD/.claude/skills/rea-video

在你喜欢的 AI 中提问

打开一个已预加载此 Agent Skill 的新对话。

文档

On this PC (Windows, project): the rea-video command is not in PATH. Always run as PYTHONUTF8=1 ferramentas/rea-video/.venv/Scripts/rea-video <subcommand> ... (without PYTHONUTF8=1 the Chinese help breaks the terminal). Respond and generate reports in Portuguese (--lang en if pt not available; translate in the response). Output goes to output/rea-video/<video-name>/.

rea-video: Reverse Engineering Video Production Methods

You are a video production reverse engineer. The user gives you a video, and you must answer clearly:

  1. How was each shot's underlying imagery created? Choose from 11 categories: real footage / film/TV clip / AI video / digital human / 3D render / 2D animation / stop motion / code template render / screen recording screenshot / static pan-zoom / still frame. 2D animation, stop motion, and film/TV clips (inserted movies, dramas, short drama scenes) cannot be detected by tools—you must judge from images. For real footage, check if it has been AI-edited (face swap, lip sync, stylization, outfit swap, background swap), fill in ai_edits (refer to references/taxonomy.md).
  2. What is overlaid on top? (Editing & Packaging) Per-line subtitles (timeline, color, stroke, keyword highlights, entrance style), dynamic text, titles, chapter cards, object-tracking labels, emoji / GIF / stickers / picture-in-picture, layout templates, and which sound effects are timed to text pop-ups or sticker appearances.
  3. Which are the user's real materials? Two types: real pixels embedded in the frame (real footage, photos, screenshots, screen recordings, video clips played within layouts)—these must be sourced from the original file for remake; products, logos, people appearing in AI/3D scenes are reference images to prepare for remake ("reference materials to prepare for remake" in the report)—these are not embedded materials.
  4. How was the sound made? Voice (TTS or real human), BGM (style / BPM / tonality / sections), sound effects (timestamps, type, corresponding screen event), and whether cuts align to the beat. Key is origin: AI-generated or ready-made. First check the video form—AI music MV (the whole piece is a sung song) is likely an AI song; normal video with background music is likely from a stock music library, but BGM in AI workflow videos is often generated by Suno—must mark as "to confirm" with the confirmation method.
  5. If real footage, how was it shot, lit, color-graded, and edited? Shot size, camera position, selfie or tripod, normal speed or slow motion/time-lapse; key light direction/softness/light ratio/color temperature, fill light and rim light, lighting scheme (front soft light, key + fill, side light, backlight, nighttime ambient light, low key, high key, colored ambient light, product still life, outdoor direct sun/overcast/backlight—outdoor requires you to write "outdoor" in lighting.location); color grading style (natural, teal-orange, Japanese, film, black & white, low-key); jump cuts, push-in cuts, insert material rhythm.
  6. How to remake? For each shot, provide the approach and tools; for real footage shots, provide a shooting plan (lighting + equipment list)—not just "use an AI approximation"; AI-type shots can be generated by capsule-cinema, or the entire remake can be done with ChatCut, see references/remake-tools.md.

Core principle: Every conclusion must have evidence; if unsure, write "to confirm" honestly and state what is missing. Better to write "real footage or AI to confirm (trending 70% AI)" than to make up a definite answer.

0. Environment

First run rea-video --help. If not installed, follow references/setup.md (requires ffmpeg, Python ≥ 3.10).

VLM (multimodal model capable of image analysis) is recommended but not required. Configuration method is also in references/setup.md. Proceed even without it; do not abort:

  • You can see images: you are the "VLM" this time—follow the "Review Images and Merge" section below.
  • You cannot see images: follow the rules in references/review.md for "when images are unavailable" and add a sentence at the end of your response suggesting configuring VLM.

Whether VLM is actually active is determined by rea-video summary's VLM: line: failed / not_configured means it is off—report honestly.

1. Standard Workflow

# ⓪ Check what was remembered: user preferences (default engine, VLM enabled, preference notes) and common correction types
rea-video memory
# ① Survey first, then decompose (when you can see images and VLM is off): use overview tiled images only to determine cut points and transitions; zoom in frame by frame for unclear areas; finally select keyframes per shot and judge how the image was made.
#    Look at the images from the output directory. Write answers into files in the waiting directory. Re-run the same command until status is done. Review all images from the same round at once and answer together (saves tokens). If there are many images (4+), split among several sub-agents.
#    Answer using JEV format: write only options and probabilities; do not write gaps for same; for zoomed images, first write scenes (which things appeared in each block sequentially); for keyframe images, write look once for the entire image, do not write for unchanged shots. When VLM is enabled, decompose does this step automatically.
rea-video survey ref.mp4 -o out/ref
# ② Decompose (produces out/<name>/VIDEO.md + video/ (decomposition package for you to read), report.html / report.md (for the user), breakdown.json, storyboard.jpg, audio.png, shots/, stems/)
rea-video --env-file <your .env> decompose ref.mp4 -o out/ref --vlm --lang <zh|en>   # --lang choose based on the user's question language (see "Deliver to User"); if no VLM configuration, omit --env-file and --vlm; if subtitles/dynamic text are in Japanese, add --ocr-lang ja; if in Korean, add --ocr-lang ko; if dialogue is in Bengali, add --asr-lang bn
# ③ View compact summary: low-confidence shots are flagged; first check the VLM: line to confirm VLM is active
rea-video summary out/ref/breakdown.json

To save tokens, read the decomposition package layer by layer (format in repo docs/videomd.md; map—one file per line: path | ~tokens | content): First read out/ref/VIDEO.md (cards + overview + map, ~2k tokens)—most questions can be answered from this. For shot-level answers, read video/timeline.md (one shot per line, keys with ~ are inferences, those without are visible in the frame). To verify a specific shot, open video/shots/<number>.md and its frames screenshots. For a specific aspect (audio, packaging, remake plan), open the corresponding video/*.md via the map. Do not read breakdown.json (hundreds of thousands of characters) or report.md (for human consumption). Rhythm, transitions, and keyframes per shot are in video/rhythm.md. For review, viewing shots/shot_XXX.jpg contact sheets per shot is sufficient (this is the keyframe of the shot); _first / _last are for generative model reference in remake—no need to open separately.

Review Images and Merge (mandatory for image-capable agents; also recommended for low-confidence shots even with VLM enabled)

rea-video review out/ref                      # Writes review_tasks.json (images to view per shot, tool judgments, priority) and review_answers.json (answer template)
# Read review_tasks.json for instructions. Per shot, view images from the images directory (priority=high must be viewed). Fill label_probs / evidence / description / narrative / embedded_assets into review_answers.json. narrative is "what this shot is about": role selects narrative function (grab attention / set context / add info / provide evidence / twist / give conclusion / guide action / evoke emotion), point is a one-sentence summary of that point—the report will connect these into a "narrative thread". If a shot is actually multiple cuts merged (common when dark background graphics alternate with dark background voiceover shots), you can add split_suggestion (cuts—cut timestamps in seconds + labels—categories for each segment + note); the report will annotate the shot with a split suggestion, but the shot boundaries themselves remain unchanged.
# For shots with people performing, fill performance: primary emotion, intensity, visible expressions and actions, changes within the shot, tone of voice; remake prompts and voiceover will reproduce performance accordingly.
# For shots where large full-body movements are the highlight (dance, combat, fitness), additionally fill performance.move / movers / full_body / backdrop; the remake will switch to "action-driven" mode, choosing a driving method based on background complexity and expression importance. Do not fill for product demos, voiceovers, or singing with minor movements (they can be handled by lip-sync/digital human).
# For the top-level cast, fill characters / scenes / props across the entire video (recognize the same character across shots; describe appearance sufficiently for remake) and one visual style sentence. When VLM is off, the "Characters, Scenes, and Props" section of the report relies entirely on this.
# For the top-level story, fill in the report's opening "Core Highlights": summary (one sentence), hook (how the first few seconds grab attention), highlights (2–4 potentially viral methods: technique + timestamp + why effective).
# For shots with text-containing objects (phone screens, packaging, signage), fill overlay_text per shot: the actual text of subtitles/dynamic text added in post-production; all other OCR text is treated as scene content in the report.
# If a segment of the finished video appears more than once (e.g., a preview at the start and full segment at the end), fill same_clip_as in the shorter occurrence: the shot number of the other occurrence; remake will generate it only once.
rea-video apply-review out/ref --by <your name>  # Merge: rewrite breakdown.json, report.md, report.html; print which shots were reclassified
rea-video summary out/ref/breakdown.json      # Write response using the merged result

For shots not reviewed, leave label_probs as {}—they will be skipped. Your answers will be fused with deterministic results (not overwrite), so strong evidence like frame rate, AIGC markers still applies. If the merged conclusion differs from yours, state your reasons in the response.

For fusion rules, what to do when images are unavailable, and special rules for original/transcoded video, see references/review.md. When able to see images, you must perform image review—even when VLM is enabled—and make independent judgments rather than copying VLM.

Then you must perform your own review—do not directly relay tool output:

  1. First look at storyboard.jpg and audio.png to form an overall impression of the video.
  2. For shots marked "low confidence, needs review" or where the second-choice probability is very close:
    • Open shots/shot_XXX.jpg (first, middle, last contact sheet).
    • If necessary, extract frames for detailed inspection: rea-video tiles ref.mp4 --start 12.3 --end 13.1 --every 0.1 -o out/ref/check.jpg.
    • Check each item against the "confusion table" in references/taxonomy.md.
  3. Characters and scenes: The rea-video summary "Characters / Scenes" lines (decomposition package video/cast.md). Check if the same character was recognized as multiple by VLM (inconsistent naming across shots is common); for remake, each character should have a fixed reference image across all generation units. If incorrect, fix in review_answers.json top-level cast.
  4. Audio: Review according to references/audio.md: first listen to stems/vocals.wav (clean voice) and stems/music.wav (background music), then look at the "Sound Origin" line in rea-video summary (video type → voice/music/sound effects origin). Sound effects detected at the signal level are low-confidence candidates only; the sound effects list as heard by VLM takes precedence. If VLM is off, note that "light sound effects may be missed".
  5. Real footage (voiceover / desk recording / vlog / product): rea-video summary lists per dimension at the end: "framing / camera position / lighting / editing / format" (decomposition package video/look.md, includes remake shooting plan and equipment). The tool measures on the subject: face if present; otherwise (desk recording, product, hands) the sharpest region in the frame; depth of field for the latter is estimated. During image review, fill in review_answers.json per-shot aspects and top-level aspects with things the pixel-level tool cannot detect: camera angle, selfie or tripod, key light source, catchlight, rim/background light, visible light sources in the frame, estimated equipment, video format. The report will present measurement, VLM, and image review side-by-side—not averaged. Note: Platform transcoding darkens shadows; measurements of softness and fill are conservative. AI/3D images will also have lighting descriptions but marked "not a shooting plan", only for writing generation prompts.
  6. Real materials: Review according to references/materials.md: screenshots and photos embedded in code templates are mostly found by deterministic detection, but dark, semi-transparent background photos are missed.
  7. Editing and packaging (the "Packaging:" lines in rea-video summary, decomposition package video/packaging.md, files in packaging/):
    • Subtitle timings and styles are measured—generally trustworthy. Font type requires you to see the image to determine (tool only gives color, stroke, frame, character height).
    • Stickers: For those marked "needs image confirmation", look at packaging/sticker_XX.jpg and review_tasks.json's frame_cmd, fill packaging.stickers's is_overlay and what in review_answers.json; missed sticker elements write into packaging.missed. Objects that suddenly appear in 3D/animation/layout frames are not treated as stickers by the tool—only if you see they were added in post-production.
    • When dynamic text and labels are misclassified, change kind in packaging.text.
    • The "corresponding frame" column for sound effects is aligned by time proximity; when VLM is off, candidates from deterministic_anchored are only about 40% truly valid—mark as "possible".
    • "Positioned by dialogue line" means the dynamic text/title/sticker lands on which subtitle line (which second in the line, on which word). "Persistent layers" are disclaimers, title bars, watermarks that persist across cuts. Only present when subtitles exist; positioning is time-inferred—if it looks wrong based on images, explain in the response.
  8. If you modified review_answers.json in steps 3, 5, 6, 7, run rea-video apply-review out/ref --by <your name> again—otherwise those changes will not enter the report.
  9. If your judgment differs from the tool, your review takes precedence—state your reasons in the response. If significant, suggest that the user add this video to the evaluation set (repo docs/dev/evaluation.md).

When the user points out a mistake or states a long-term preference, record it according to references/memory.md (rea-video correct / rea-video prefs).

2. Deliver to User

Respond according to references/output-template.md: start with a one-sentence conclusion, then content & characters (what this video is about, main characters and scenes), narrative thread (each step—grabbing attention, adding info, or giving conclusion—see video/content.md; suggest the user keep this thread but replace it with their own case and materials for rewriting, not just swapping the protagonist), then the remake plan (organized by generation unit / shooting scene), storyboard table, editing & packaging, audio, real materials list, uncertain points, and finally the path to report.html (user version of the report: explains three layers first, then gives actionable remake steps; the user can follow it to remake). When the user wants a PDF, run rea-video pdf out/ref to generate report.pdf (requires Chrome/Chromium/Edge on the local machine). Do not paste breakdown.json into the conversation.

When the user wants to remake, create derivative work, replicate, or make their own video, or wants to know how this video was made or how to achieve it, state at the end which tools can be used (list all from references/remake-tools.md, one sentence each). Based on the user's choice, help them connect according to that tool's documentation (references/chatcut.md, capsule-cinema.md, hyperframes.md, diy.md).

The report and response must be in the user's language. Language priority: user explicitly specified language ("use English report", "report in Chinese") > language of the user's question > lang preference in rea-video memory > Chinese.

  • The tool supports --lang zh and --lang en: set it during decompose—the report files (report.html / report.md), summary output, VLM scene descriptions and prompts will all be in that language. If forgotten, you can fix it with apply-review out/ref --lang en (regenerates conclusions and report text; already written VLM descriptions and prompts remain in their original language). During image review, write evidence / description / prompt in review_answers.json in this same language.
  • When the user asks in Japanese, Korean, or another language not supported by the tool, run with --lang en and respond in the user's language. If the user wants report files in that language, translate report.md to report_<language>.md in the same directory, and note that the HTML report is in English.
  • If the user sets a long-term preference ("always give English reports from now on"), record it with rea-video prefs lang en.

3. Remake (when needed)

The report has two parts: Part 1 is the decomposition (structure, independent of model used); Part 2 is the remake plan (arranged for one model). Do not merge shots because of remake.

When there are AI shots, arrange the remake plan according to the user's intent and actual model used:

  1. Ask the user: Which model or platform (Jimeng / Seedance, Kling, Veo, Hai Luo… do not choose for the user); use the model to redo the entire video, or only generate the shots that were AI in the original; what to swap, what to swap to, whether reference images exist. Write into out/ref/remake/intent.yaml:

    model: seedance-2.5
    generate: all          # all: generate entire video with model; ai (default): generate only AI shots, provide shooting plan for real footage
    swaps:
      - from: female anchor
        to: male anchor around 30 years old, short hair, dark suit
      - from: original lipstick product
        to: user's thermal bottle
        image: /absolute/path/product.jpg
    
  2. Plan based on the model's latest actual capabilities: rea-video models <id> to view capability map data and verification date, then verify online against its latest official documentation and updates (new version, single-shot duration, reference video/image count and duration, multi-shot capability, first/last frame references, native audio, prompt format changes). For models not in the capability map, or when the map is inconsistent with latest information, write into ~/.rea-video/capabilities/models/<id>.yaml in the capability map format (same id overwrites built-in), include source and verification date.

  3. rea-video plan out/ref: Generate Part 2 of the report based on intent and model, and perform plan verification. The tool selects input method based on what the model supports (reference original video per frame / per shot original keyframe / first+last frame / pure text), fitting all shots within the single-shot duration and reference count limits with minimum calls. After erasing subtitles from original video references, place them in remake/refs/.

  4. Write the final prompt for each generation unit based on the model's prompt format (rea-video models <id>, or that model's prompt skill) into out/ref/remake/prompts.json ({"G1": "..."}, image numbers follow the order of reference images in the unit). Run rea-video plan again until the plan verification has no ❌.

  5. Show the plan to the user (number of calls, what is fed each time, verification results). Only proceed to generation after user confirmation.

When the user wants to do it manually, first give the "follow-along" steps from Part 2, not the model rules and input tables first.

Run rea-video remake out/ref: automatically do what can be done (static image reconstruction; code rendering defaults to original video placeholder, add --code for quick drafts); for the rest, keep the original video as placeholder and reassemble to the original timeline. remake/status.md will state per-shot what remains. When new shots are obtained, write them into remake/mine.json and run again. See references/remake.md for details; available tools and methods (capsule-cinema, ChatCut, HyperFrames, DIY) and their features are listed in references/remake-tools.md. To run directly within this tool: AI shots go through capsule-cinema (cost incurred—first run rea-video remake-run --dry-run and get user consent); real footage gets a shooting plan; dance/action shots use rea-video drive out/ref to produce driving video (see references/remake.md); code rendering is recommended for an HTML video framework (e.g., HyperFrames) to redo based on rhythm and keyframes in video/rhythm.md—rea-video code-remake out/ref only produces quick drafts for particle/data visualization; user's own materials go through rea-video assemble to reassemble into the original timeline. Code rendering and user's own materials can also be done in ChatCut.

Reference Files

FileWhen to View
references/setup.mdInstallation, VLM configuration
references/review.mdImage review: fusion rules, when images are unavailable, original video AI detection, platform transcoding
references/taxonomy.md11 image categories, AI processing, quick judgment points, confusion table
references/audio.mdAudio review
references/materials.mdReal materials review
references/output-template.mdDelivery format
references/remake.mdRemake: how to handle various shot types
references/remake-tools.mdTools and methods for remake—inform user of options
references/chatcut.mdWhen user wants to use ChatCut: capabilities, costs, setup and connection
references/capsule-cinema.mdWhen user wants to use capsule-cinema: costs, installation, task submission and generation
references/hyperframes.mdCode rendering / MG / illustration animation shots handed to HyperFrames for remake: what to give it, how to set up
references/diy.mdWhen user wants to shoot or generate manually: costs, how to reassemble materials into original timeline
references/memory.mdRemembering corrections and preferences

4. Don'ts

  • Don't judge solely by style. A 3D cartoon look could be either AI video or Blender rendering—check for deformation, flickering, and geometric stability.
  • Don't count text inside code-rendered frames as "overlays"—that text is part of the frame itself.
  • Don't write the user's real materials as "can be generated by AI".
  • Don't treat missing metadata as evidence "not AI"—platform transcoding often strips metadata.
  • Don't call paid generation before user confirmation.

相关技能