抖音视频分析

Overview

Turn a Douyin video link into an evidence-backed research report. Prefer parsed page copy, captions, local/free transcription, and key frames over unsupported inference.

Workflow

1. Resolve the input
   • Extract the first URL from share text.
   • Record the original URL, resolved URL, redirect chain, status code, and crawl method.
   • If a compatible parser exists in the active workspace, prefer its documented CLI over ad hoc scraping. This Skill does not bundle a parser runtime. For a workspace that provides the douyin_analyzer package, a compatible example is: python -m douyin_analyzer.cli analyze "<share text or URL>" --out "reports/douyin-video-analysis/<run-id>" --stt-backend auto
2. Collect evidence
   • Extract page title, author, description, hashtags, captions, and video/media URLs.
   • If HTTP parsing fails, use browser rendering/network capture and save diagnostics.
   • Download media only when the source URL is accessible and allowed by the current session constraints.
   • Use free/local speech-to-text first: page captions, then faster-whisper, then FunASR; do not use paid STT unless the user explicitly asks.
   • Extract 3-7 key frames across the video or around uncertain transcript sections.
3. Separate evidence from interpretation
   • Treat transcript text, page metadata, frames, and network diagnostics as evidence.
   • Mark low-confidence ASR, missing captions, blocked pages, or inferred claims.
   • Never invent exact quotes, figures, interactions, claims, or visual details not present in evidence.
4. Analyze deeply
   • Identify the video's thesis, argument structure, assumptions, examples, and implied framework.
   • Compare page copy with transcript and frames; flag mismatches.
   • For technical, financial, or trading videos, distinguish education, strategy rules, unverifiable claims, and risk warnings.
   • Use key frames to adjudicate ambiguous transcript passages, especially charts, slides, formulas, or UI demonstrations.
5. Write the report
   • Use the structure in references/report-framework.md.
   • Make the Markdown report a finished learning/research report, not a transcript dump or artifact index.
   • Use transcript and frames as evidence, but do not paste the full transcript or intermediate artifact paths into the main report unless the user explicitly asks.
   • Explain terms, techniques, principles, what each technique is used for, validation boundaries, and what the viewer can learn.
   • Keep raw transcript, media paths, diagnostics, and key frame paths in machine-readable artifacts such as JSON or a separate appendix when available.

Failure Handling

• If crawling fails, do not stop at "failed". Save and analyze the page structure: HTML snapshot, scripts summary, HTTP metadata, browser screenshot, and network log.
• If media downloads fail but metadata exists, produce a partial report and list the missing evidence.
• If transcription is unavailable, explain the missing free backend briefly and still analyze page copy and frames.

Quality Bar

• A report is not complete until it reads like a useful study note: source summary, content structure, terminology, techniques/principles, learning path, uncertainties, and validation boundaries.
• Raw transcript and intermediate artifact paths are evidence storage, not the body of the Markdown report.
• Re-run the narrowest relevant command before claiming the result is complete.
• Keep the report useful for a reader who will not inspect the raw video.