video-performance-analyzer（视频性能分析仪）

Video Performance Analyzer

What This Skill Does

Takes a short-form video (MP4 file or YouTube URL) and returns three things:

1. Full transcript — every spoken word, on-screen text, audio cues, with timestamps
2. Performance analysis — why it works, mapped against proven growth frameworks
3. Repurposing playbook — concrete content ideas derived from the video

Setup

The analysis script uses Google's Gemini API (generative model for transcript + analysis).

API Key: Read from .env at repo root — GEMINI_API_KEY

Install dependency (one-time):

pip install google-genai --break-system-packages

Script location: <skill-directory>/scripts/analyze_video.py

Input Formats

Running the Analysis

Quick run

python <skill-directory>/scripts/analyze_video.py /path/to/video.mp4

YouTube URL

python <skill-directory>/scripts/analyze_video.py "https://www.youtube.com/watch?v=VIDEO_ID"

With output saved to file

python <skill-directory>/scripts/analyze_video.py /path/to/video.mp4 --output /path/to/output.md

The script prints structured Markdown to stdout. Pipe or redirect as needed.

Analysis Workflow

Step 1: Run the script

Run analyze_video.py with the video path. The script will:

• Upload the video to Gemini Files API (or pass YouTube URL directly)
• Wait for processing (usually 10-30 seconds)
• Send the analysis prompt to gemini-3-flash-preview
• Return raw structured output

Step 2: Map against growth frameworks

After the script returns, interpret the output through these growth dimensions:

Hook — "The Hook Is a Contract"

• Did the hook make the viewer feel implicated, called out, or urgently curious?
• Does it pass any of the 5 hook techniques? (Contradiction, Specific number + unexpected context, Direct accusation, Stolen thought, Absurd reframe)
• Is it about the viewer's situation — or the creator's achievement?
• Framing trick to look for: deliberate word choice that raises perceived stakes

Structure — Progressive Intrigue

• Map each timestamp beat and score whether stakes escalate beat-to-beat
• Look for the "therefore/but" chain — does each beat cause the next, or just follow it?
• Identify where the information gaps open and where they close
• Flag any beat that could be predicted from the one before it

Credibility — Specificity Signal

• Count concrete numbers, named people, specific dates, dollar amounts
• Note any "documentary proof" moments (showing receipts/sources visually as claims are made)
• Generic claims = weak. Named + numbered claims = strong.

Framing — Reader-First

• Is the video about the creator's achievement, or the viewer's situation?
• Every strong performing video frames its value around what the *viewer* gets
• Personal milestone framing = performs with existing audience, dies everywhere else

Retention — Visual Pattern Interrupts

• How often does the visual change? Every 5-10s is the target
• Better backgrounds / overlays = higher retention = platform pushes to new audiences
• Note any transitions, overlays, text cards, B-roll cuts that reset attention

CTA — Loop-Closing vs. Loop-Opening

• Does the close resolve the viewer's question, or leave one open?
• Best CTAs create a condition where following is the obvious next move (implied CTA)
• Weak CTAs: "like and follow if you learned something", "let me know in the comments"
• Strong CTAs: "we'll see what happens in a few hours" (creates forward momentum)

Step 3: Generate repurposing ideas

Based on the analysis, identify every angle that could become standalone content. For each idea:

• What format fits best (short-form video, LinkedIn post, X thread, newsletter section)
• What the hook/banner should be
• Whether it maps to a content pillar (AI Implementation Reality, Behind the Scenes, Industry Hot Takes, Democratizing Expertise)
• Funnel stage (TOFU / MOFU / BOFU)

Repurposing sources to mine:

• The strongest single claim in the video -> standalone hot take post
• The "before/after" or "X vs. Y" implicit in the narrative -> comparison format
• Any step-by-step workflow mentioned -> listicle or tools video
• The emotional core (fear, surprise, outrage, relief) -> personal angle script
• Any stat or number -> data-driven short-form
• The contrarian framing (if present) -> X thread or newsletter hook

Step 4: (Optional) Push to your content calendar

If you use Notion as a content calendar, you can push repurposing ideas after analysis.

Configure your Notion data source ID in your project settings:

data_source_id: <your-notion-database-id>

Default fields for repurposed content:

• Status: Backlog
• Platform: based on format fit
• Content Format: ["Video"] for scripts, ["Text"] for posts
• Content Type: based on angle
• Funnel Stage: TOFU (most repurposed content is awareness-level)
• Re-purpose?: Yes

Output Format

Deliver results in this order:

1. Full Transcript

Timestamped, verbatim. Include on-screen text overlays noted inline.

2. Visual & Audio Breakdown

• Format (talking head / voiceover / B-roll / screen recording)
• Background / set design quality
• Music / audio tone
• Visual change frequency (pattern interrupts per minute)

3. Performance Analysis

Score each dimension (Strong / Moderate / Weak):

Then 3-5 bullet points on *why* this video performs — the specific mechanisms, not generic observations.

4. Repurposing Playbook

For each idea:

**[Format] — [Working title / angle]**
- Hook: ...
- Format: ...
- Content pillar: ...
- Funnel stage: ...
- Why it'll work: 1 sentence

Aim for 4-8 repurposing ideas per video. At least 2 should be immediately scriptable.

5. What to Steal

3-5 specific, actionable techniques from this video that you should copy directly into your own content system. Concrete, not generic. Reference the specific moment in the video where each technique appears.

What This Skill Does NOT Do

• Download TikTok or Instagram videos (download manually, then pass the path)
• Post or publish anything — analysis and ideation only
• Access private or unlisted YouTube videos
• Replace your creative judgment — the analysis surfaces what's working, your taste picks what to steal

Notes on Model Selection

• Transcript + analysis: gemini-3-flash-preview — best video understanding, handles 3-5min short-form easily
• Embedding (future): gemini-embedding-2-preview — for comparing videos semantically against a library of performers. Not yet implemented in the script but the architecture is ready.
• Free tier limits: 8 hours of YouTube video/day. No limit on uploaded files (paid tier). Short-form videos are well within limits.