video-structure-analyzer
Skill4 starsChanged 5 months agoMITView on GitHub
拆解爆款短视频、Vlog、小红书/抖音/B站视频、广告素材或竞品视频时使用。用于把视频链接、截图、录屏、字幕、逐字稿或参考视频拆成镜头级拉片表、爆款机制、可复用结构、批量选题变量、脚本/分镜/prompt,以及给后续视频生成 agent 使用的 JSON brief。也适用于用户说“分析爆款视频”“拆视频”“拉片”“拆解短视频”“复刻这种视频结构但不要抄”“根据爆款批量产出视频方案”等场景。
1 other project uses a version of this, 1 of them edited.
Adopted by
Most recently taken up by
- F0rger123/claude-video-editing-god-modeedited, first seen Nov 1, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 1
- F0rger123/claude-video-editing-god-mode★ 182% same
Wrote their own version · 0
None found.
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repositoryOne other file it gives its agents.
The file
Show its rules and text
We split this skill into the rules it gives an agent
23 rules. About 0 are common, 12 are distinctive, and 11 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈12
Distinctive · estimate
This author's own choices. Worth a look.
11
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 10
Use this skill as the pre-production researcher for video-generation agents.
- Very specific: line 18
When the user asks for an HTML report, the report is not a free-design task.
- Very specific: line 22
Keep the interface name Viral Video Lab.
- Very specific: line 23
Keep the same A/B switch behavior, labels, navigation, module order, spacing scale, card treatment, table treatment, shot-card grid, mobile …
- Distinctive: line 24
Do not invent a new hero, color palette, card style, Chinese title bar, marketing landing page, decorative gradients, special blueprint …
- Distinctive: line 55
If a platform blocks direct download, the skill should:
- Distinctive: line 58
Use screenshots or metadata as alternatives
- Very specific: line 67
Automatic video download: Use yt-dlp, ffmpeg, or browser automation to automatically download the video from the provided URL.
- Very specific: line 71
If blocked, try ffmpeg with URL directly
- Distinctive: line 72
If still blocked, use browser automation to capture video
- Distinctive: line 77
Always record the extraction method used in the source facts.
- Very specific: line 93
Create one S### row/card per visible cut or meaningful continuous action beat.
- Distinctive: line 97
Every shot card in the HTML should have a matching real image file.
- Distinctive: line 98
If exact frames are unavailable, do not invent precise timecodes.
- Very specific: line 133
Do not include a visible JSON Brief(机器可读) module inside the HTML report unless the user explicitly asks to display raw JSON in the web page.
- Distinctive: line 141
Use this table unless the user asks for another format:
- Distinctive: line 190
Treat empty shots as meaningful: B-roll, pauses, object close-ups, and transition shots often carry rhythm and emotion.
- Distinctive: line 192
If evidence is incomplete, mark confidence and avoid pretending exact frame timing is known.
- Distinctive: line 193
Automatic video handling: Always attempt to download video from URL automatically.
- Very specific: line 194
Before delivering an HTML report, verify that the HTML contains: .topbar-inner, .brand with Viral Video Lab, .style-switch, …
- Very specific: line 195
Before delivering an HTML report, verify that every local <img src="..."> path resolves relative to the HTML file.
- Very specific: line 196
Before delivering an HTML report, verify that it does not contain placeholder-only shot panels, a visible JSON Brief(机器可读) section, or …
- Very specific: line 197
If possible, run references/validate_report.py <report.html> before final delivery.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Video Structure Analyzer
Use this skill as the pre-production researcher for video-generation agents. The goal is not only to explain why a video works, but to produce reusable, original, machine-readable production specs.
Core Rule
Analyze methods, structure, rhythm, and production logic. Do not copy protected expression. Avoid reproducing exact scripts, distinctive shot sequences, watermarks, creator identity, brand identity, or proprietary assets unless the user owns them or explicitly asks for analysis-only commentary.
Non-Negotiable Report Fidelity
When the user asks for an HTML report, the report is not a free-design task. You MUST reproduce the accepted Viral Video Lab report system exactly enough that a user can recognize it as the same product:
- Load
references/report-style.mdbefore writing HTML. - Inline or copy the exact CSS from
references/viral-video-lab.css; do not replace it with a new visual system. - Keep the interface name
Viral Video Lab. - Keep the same A/B switch behavior, labels, navigation, module order, spacing scale, card treatment, table treatment, shot-card grid, mobile topbar behavior, and
AI 生产蓝图module styling. - Do not invent a new hero, color palette, card style, Chinese title bar, marketing landing page, decorative gradients, special blueprint panel, or unrelated template.
- Content may change per video; layout and visual system should remain stable.
Input Modes
Primary mode: Video URL only - Users should only need to provide a video link. The skill automatically handles video download and processing.
Accepted video sources:
- Xiaohongshu (小红书): Direct link or video ID
- Douyin (抖音): Direct link or video ID
- Bilibili: Direct link or video ID
- YouTube: Direct link or video ID
- TikTok: Direct link or video ID
- Kuaishou (快手): Direct link or video ID
- Instagram: Direct link or video ID
- Ad libraries: Direct link
Alternative inputs (when video link is unavailable):
- Local video file upload (for testing or when link is blocked)
- Screenshots or screen recordings
- Transcript, SRT/VTT subtitles, or OCR text
- Existing analysis notes
Important: Users should NOT need to manually download or upload videos. The skill should automatically:
- Accept the video URL
- Download the video using appropriate tools (yt-dlp, ffmpeg, browser automation, etc.)
- Extract frames and metadata
- Generate the analysis report
If a platform blocks direct download, the skill should:
- Attempt multiple download methods
- Fall back to browser automation if needed
- Use screenshots or metadata as alternatives
- Clearly state any limitations in the output
Workflow
- Capture source facts: Title, creator/account, platform, URL, duration, visible engagement metrics, publish date if available, product/category, and any extraction limitations.
Extract usable evidence: Automatic video download: Use
yt-dlp,ffmpeg, or browser automation to automatically download the video from the provided URL. Do not ask the user to upload the video file.Extraction priority:
- Attempt automatic download using
yt-dlp(supports most platforms) - If blocked, try
ffmpegwith URL directly - If still blocked, use browser automation to capture video
- Extract frames using
ffmpegorffprobe - Extract subtitles/transcripts if available
- For blocked web video, sample screenshots at meaningful timestamps
Always record the extraction method used in the source facts.
For full HTML reports, evidence images are required. Create a sibling visual directory for every report, named with the same slug plus
-visual, for example:xhs-smelly-cat_图文拆解报告.htmlxhs-smelly-cat-visual/contact_sheet.jpgxhs-smelly-cat-visual/S001.pngxhs-smelly-cat-visual/S002.png
The HTML must reference those local relative files. A final
*_图文拆解报告.htmlmust use real frame images, not decorative placeholders. If frames cannot be extracted, do not pretend the report is complete: create a clearly labeled draft analysis or ask the user for the video file/screenshots, and state that a final HTML report requires visual evidence. Do not silently omit the screenshot/contact-sheet modules.- Attempt automatic download using
Build the shot-level table: The table is the main output. Include timecode, duration, visual, shot size, action, subtitle/dialogue summary, sound, camera angle, camera movement, transition, emotional function, narrative function, and generation prompt.
For full reports,
Shot-by-Shotmeans actual shots/cuts, not broad narrative chunks. Do not compress a long video into 6-8 generic sections unless the user explicitly asks for a compressed outline. When real frames are available:- Create one
S###row/card per visible cut or meaningful continuous action beat. - For slow continuous footage, sample at least every 3-5 seconds.
- For 20-60s videos, expect roughly 10-25 shots unless the video is truly static.
- For 60-120s videos, expect roughly 18-45 shots unless the video is truly static.
- Every shot card in the HTML should have a matching real image file.
- If exact frames are unavailable, do not invent precise timecodes; label the section as a reconstructed outline instead of final shot-by-shot.
- Create one
- Segment the narrative: Group shots into sections such as Hook, setup, proof, reveal, contrast, routine block, escalation, payoff, summary, CTA. Name the section by its job, not by vague labels.
- Explain the viral mechanism: Identify hook type, curiosity gap, emotional driver, information density, rhythm, visual density, trust signals, audience fantasy/pain, repeatable meme or comment trigger, and conversion move.
- Convert analysis into production assets:
Produce reusable templates, variables, batch variants, script options, shot prompts, caption rules, BGM/SFX guidance, platform adaptations, and a
video_generation_brief.json-style object. - Add originality guardrails: List what can be reused at the pattern level and what must be changed to avoid copying.
Required Outputs
For quick requests, answer inline with these sections:
- Source facts
- One-paragraph thesis
- Shot-level table or compressed shot table
- Viral mechanism
- Reusable production template
- Batch variants or next-video ideas
- JSON brief
For substantial requests, create files in the working project:
*_拆解报告.htmlor*_拆解报告.md*_拉片拆解.md*_video_generation_brief.json- Required for HTML reports: a
*-visual/directory containingcontact_sheet.jpgplus sampled frame images namedS001.png,S002.png, etc. Usevisual/contact_sheet.jpgpaths only when the HTML itself lives inside the same parent directory and the folder is actually namedvisual.
When producing HTML reports, read references/report-style.md and follow that style contract as a hard requirement. Prefer a dense professional analysis surface, not a marketing landing page. The accepted default is a Viral Video Lab report with source facts, stats, contact sheet, shot cards, production table, AI blueprint, variable slots, prompt template, and an A/B visual style toggle.
Do not include a visible JSON Brief(机器可读) module inside the HTML report unless the user explicitly asks to display raw JSON in the web page. The JSON brief belongs in the separate *_video_generation_brief.json file. The HTML should include Agent Prompt Template, not raw machine JSON.
The AI 生产蓝图 module should use the same module/card styling as adjacent sections in both A and B themes unless the user explicitly asks for a special treatment.
For key-shot overviews, generate a compact contact sheet with tight gutters. Do not use large screenshot row gaps that make the report excessively long.
Shot Table Schema
Use this table unless the user asks for another format:
| 排序 | 镜号 | 时间码 | 时长秒 | 画面 | 景别 | 内容/动作 | 台词/字幕 | 声音 | 摄像机角度 | 镜头运动 | 转场 | 情绪功能 | 叙事功能 | 生成提示词 |
For fast analysis, compress to:
| 镜号 | 时间码 | 画面/动作 | 台词/字幕摘要 | 情绪功能 | 叙事功能 | 可复用生成提示 |
JSON Brief
For the complete JSON contract and detailed heuristics, read references/output-contract.md when generating machine-readable assets.
Minimum JSON object:
{
"source": {},
"analysis_summary": "",
"viral_pattern": {},
"shots": [],
"production_template": {},
"batch_variants": [],
"originality_guardrails": []
}User Experience
Minimal user input required: Users should only provide:
- Video URL (or video ID)
- (Optional) Specific analysis focus or output format
Automatic handling:
- Video download
- Frame extraction
- Subtitle/transcript retrieval
- Metadata collection
- Report generation
No file upload needed unless:
- The video URL is blocked and user provides local file as fallback
- User wants to analyze multiple videos at once
Quality Bar
- Be concrete: prefer "0-3s close-up of messy desk, abrupt zoom, anxiety hook" over "strong opening".
- Preserve timing: timecodes and durations matter because later agents need executable pacing.
- Treat empty shots as meaningful: B-roll, pauses, object close-ups, and transition shots often carry rhythm and emotion.
- Separate human-readable insight from machine-readable production specs.
- If evidence is incomplete, mark confidence and avoid pretending exact frame timing is known.
- Automatic video handling: Always attempt to download video from URL automatically. Only ask for file upload if automatic download fails after all retry methods.
- Before delivering an HTML report, verify that the HTML contains:
.topbar-inner,.brandwithViral Video Lab,.style-switch,body[data-style="swiss"],#overview,#shots,#blueprint,#prompt, a contact-sheet<img>, and at least one shot card image when frames are available. - Before delivering an HTML report, verify that every local
<img src="...">path resolves relative to the HTML file. Broken image references are a failed output. - Before delivering an HTML report, verify that it does not contain placeholder-only shot panels, a visible
JSON Brief(机器可读)section, or compressed narrative beats mislabeled as final shot-by-shot analysis. - If possible, run
references/validate_report.py <report.html>before final delivery. If it fails, fix the report instead of presenting it as complete.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

