功能描述

抖音视频分析流水线：输入抖音链接后，自动/半自动抓取页面信息与音频、转写口播、总结内容、提取有用信息与对应证据、进行批判性思维分析，并把转写与分析结果写入 Obsidian。 / Analyze a Douyin video from a shared URL by extracting page metadat...

使用说明 (SKILL.md)

Douyin Video Analysis

Name: Douyin Video Analysis
Author: zzmbobby

Use this skill when the user gives a Douyin video URL and wants:

the real video topic/content extracted
the title and hook analyzed
audio downloaded from the page if possible
audio transcribed into readable Chinese text
critical thinking analysis of the script/content
results saved into Obsidian notes

Scope

This skill is for analysis and note capture, not reposting or bulk scraping.

Inputs

Expected user input:

a Douyin short URL or full video URL
optional analysis focus, such as:
- whether the copy is AI-generated
- title strategy
- persuasion methods
- logical flaws / manipulation / exaggeration

Output contract

Produce:

short status update in chat
transcript note in Obsidian
analysis note in Obsidian
final chat summary with note paths

Workflow

1. Resolve the real video page

Open the Douyin URL in Chrome if needed.
Extract:
- canonical video URL
- title
- description
- account name
- visible chapter points / page text
Be explicit about what came from page metadata vs full transcript.

2. Extract media URLs if possible

Prefer reading browser performance resources from the actual loaded page.
Capture candidate audio/video stream URLs.
If direct download gets 403, retry with browser-derived headers/cookies when available.
Save temporary files under /tmp/douyin_transcribe/.

3. Transcribe audio

Prefer a local/offline path.
Current practical path on this machine:
- Python venv in /tmp/douyin_transcribe/venv
- mlx-whisper
First pass:
- use a smaller model to get a usable draft quickly
If needed:
- do a second pass or manual cleanup for readability
Clearly label transcript status:
- raw machine transcript
- cleaned readable version
- final polished version

4. Summarize before criticizing

Before any critical-thinking judgment, always do these two steps:

A. Content summary

produce a concise summary of what the video actually says
separate hook / core claim / support / conclusion
avoid mixing your judgment into the summary

B. Useful information + evidence extraction

List the video's actually useful information and pair each item with the evidence the video itself provides. For each item, give an initial confidence judgment:

high
medium
low

This prevents the analysis from becoming pure attitude.

5. Analyze critically

Always cover, unless the user narrows scope:

what the video is actually saying
whether the title/hook exaggerates or distorts the case
whether the script appears AI-generated / AI-assisted / human-written
emotional triggers used
hidden assumptions
survivorship bias / overgeneralization / false lessons
what is genuinely useful vs what is manipulative or shallow

6. Save into Obsidian

Default destination:

transcript note → OpenClaw/Inbox/YYYY-MM-DD 抖音视频口播转写 - \x3Ctopic>.md
analysis note → OpenClaw/Inbox/YYYY-MM-DD 抖音视频分析 - \x3Ctopic>.md

Include:

source URL
canonical URL
account name if available
what was directly obtained vs inferred
transcript quality warning if imperfect

Response rules

Never pretend a full transcript was obtained if only page metadata was captured.
If only metadata is available, say so plainly.
Distinguish between:
- page title/description analysis
- full script analysis
Prefer usefulness over theatrical certainty.

Local implementation notes

Useful browser path

Probe browser bridge
Use macOS Chrome bridge to:
- open the page
- evaluate page JS
- inspect visible text
- inspect performance.getEntriesByType("resource")

Useful temporary files

/tmp/douyin_transcribe/audio*.mp4
/tmp/douyin_transcribe/venv

Suggested helper scripts

A. Grab page and audio

python3 /Users/bobzhong/.openclaw/workspace/skills/douyin-video-analysis/helpers/douyin_grab.py '\x3Cdouyin-url>'

It will:

open the Douyin page in Chrome
fetch page metadata from the loaded Chrome tab
extract candidate media URLs
attempt authenticated audio download via curl with copied headers/cookies
return JSON containing canonical URL, title, body preview, audio URL, and local audio path

B. Transcribe and create notes

python3 /Users/bobzhong/.openclaw/workspace/skills/douyin-video-analysis/helpers/transcribe_and_note.py \
  --audio-file /tmp/douyin_transcribe/audio_latest.mp4 \
  --title '\x3Cvideo-title>' \
  --source-url '\x3Cshared-url>' \
  --page-url '\x3Ccanonical-url>' \
  --account '\x3Ccreator>' \
  --topic '\x3Cshort-topic>'

It will:

transcribe the local audio with mlx-whisper
create a transcript note in Obsidian
create an analysis-note skeleton in Obsidian
return both note paths as JSON

Limits

Some videos may block direct media download.
Machine transcripts may contain proper noun and number errors.
Large-model transcription may be slow; start fast, then refine if needed.

When done

Tell the user:

whether transcript is metadata-only / rough transcript / cleaned transcript
where the notes were saved
whether further cleanup would materially improve the analysis

安全使用建议

What to check before installing/running: - Paths and environment: the scripts use hard-coded paths (/Users/bobzhong/…, chrome bridge path, and an iCloud/Obsidian vault). If you're not the original author, update these paths to match your machine to avoid accidental writes or failures. - Browser bridge and cookies: the skill uses a Chrome bridge to run JS and reads document.cookie, then includes that cookie in curl requests to download audio. This is necessary to fetch some protected Douyin media but exposes session cookies. Only run if you trust the chrome bridge script and are comfortable allowing the skill to use the page's cookies. If you don't want to expose your session, run the capture in a logged-out/incognito profile or remove the cookie use. - Network activity: transcription via mlx-whisper may download models from Hugging Face or other repos at runtime. Expect outbound network traffic for model retrieval. - Inspect helper scripts locally: the included Python helpers invoke curl and subprocesses; they appear coherent for the stated purpose and don't contain obfuscated endpoints, but review them and the chrome bridge implementation before running on a sensitive machine. - Permissions and data: the skill will write notes into your Obsidian vault (local/icloud path). Make sure you are okay with analyzed content (and any downloaded audio) being stored there. If you want to proceed safely: update the hard-coded paths to configurable values, verify the chrome bridge script, run first on a throwaway account or in an isolated profile, or modify the code to avoid sending cookies if you only need public metadata.

功能分析

Type: OpenClaw Skill Name: douyin-video-analysis Version: 0.1.4 The skill bundle implements a complex pipeline for scraping Douyin, downloading audio, and transcribing it. While the behavior aligns with the stated purpose, it performs several high-risk actions: it extracts active browser cookies and User-Agent strings via a Chrome bridge (`douyin_grab.py`), uses `subprocess.run` to execute shell commands with web-derived data, and relies on hardcoded absolute paths for a specific user profile (`/Users/bobzhong/`). These capabilities, particularly the programmatic extraction of browser cookies and the potential for command injection, represent significant security and privacy risks that warrant a suspicious classification despite the lack of clear evidence of intentional malice.

能力评估

ℹ Purpose & Capability

The skill's name/description (Douyin analysis → fetch metadata/audio, transcribe, analyze, save to Obsidian) matches what the code and SKILL.md do. It legitimately requires a browser bridge and local Whisper transcription to access protected audio and create local notes. However, the implementation hard-codes user-specific paths (/Users/bobzhong/... and iCloud Obsidian vault path), which reduces portability and indicates the package is tailored to a particular machine rather than being broadly generic.

ℹ Instruction Scope

Runtime instructions explicitly say to open Chrome, inspect performance resources, capture candidate media URLs, and use browser-derived headers/cookies to download audio; the code does exactly that (chrome bridge eval, document.cookie, performance entries). Accessing page metadata, resources, and writing transcripts/notes to Obsidian are in-scope. The cookie capture and use for curl are sensitive operations but are declared in SKILL.md.

✓ Install Mechanism

There is no install spec (instruction-only), and the repository contains helper scripts executed at runtime. No downloads or third-party archives are fetched by the skill itself. The skill assumes a temporary Python venv where mlx-whisper can be installed; actual venv creation/installation is left to the runtime environment. It relies on an external chrome bridge script (path assumed), which is a subtle external dependency to verify.

ℹ Credentials

The skill requests no declared environment variables or credentials, which matches the registry metadata. It does, however, read document.cookie from the opened page and uses that cookie in curl requests to retrieve protected media — this is necessary to access authenticated Douyin media but involves sensitive session tokens. The code also writes to a user-specific Obsidian vault path and expects macOS + Chrome + browser bridge. These accesses are proportionate to the stated goal, but they are privacy-sensitive and environment-specific.

✓ Persistence & Privilege

The skill does not request always:true or elevated platform privileges. It writes temporary files under /tmp/douyin_transcribe/ and writes notes into the user's Obsidian Inbox (hard-coded path). Writing to the user's own notes and the temp directory is consistent with the stated behavior and not an inherent privilege escalation.

版本历史

v0.1.4

Make the README visually Chinese-first by moving Chinese titles and descriptions ahead of English throughout the page.

v0.1.3

Add stronger Chinese-first description and move Chinese listing copy ahead of English for ClawHub-facing docs.

v0.1.2

Add stronger bilingual listing copy and rewrite README as a more complete product-style page.

v0.1.1

Add bilingual docs, fuller capability guide, usage instructions, outputs, limits, and ClawHub listing notes.

v0.1.0

Initial release: capture Douyin metadata/audio, transcribe, summarize, critically analyze, and save into Obsidian.

元数据

Slug douyin-video-analysis

版本 0.1.4

许可证 MIT-0

累计安装 7

当前安装数 7

历史版本数 5

常见问题