SuperAPI AI Directory
Partner中文AI newsFavorites
Skillssqsl-video-to-article

sqsl-video-to-article

Design & creation

View repository

Updated 10-16 09:55

About this skill

🎬 SQSL Video To Article (Video to Illustrated Article Generation Engine)

Quickly transform local video materials intoillustrated, highly picture-aligned, restrainedly polished, 100% preserving the author's original wording and stylehigh-quality Markdown article draft.

[!NOTE] This skill focuses on the core production stage of **"video material ➔ subtitle correction ➔ precise frame extraction ➔ restrained original illustrated Markdown"**. The resulting Markdown manuscript can be directly handed to the family typesetting engine sqsl-article-to-wechat to render into a WeChat official account magazine-style HTML and push directly to the draft box with one click.

Core Process (4-Step Closed Loop)

flowchart TD
    A[输入本地视频 .mp4 / .mov] --> B[1. ffmpeg 提取 16kHz mono WAV 音频]
    B --> C[2. 本地 whisper-cli 极速转写为 SRT / TXT]
    C --> D[3. 常用词典自动纠错 references/glossary.json]
    D --> E[4. 基于 SRT 时间戳精准抽帧与回看校验]
    E --> F[5. 撰写 Markdown：极度克制改写，保留作者原话，仅剔除口水词，零 AI 味]
    F --> G[🎯 下一步：交由 sqsl-article-to-wechat 杂志排版与直推]

Step 1: Local Audio Extraction and Ultra-Fast Transcription

No need to upload huge videos to cloud models, directly use local ffmpeg + whisper-cli(equipped with Metal / GPU hardware acceleration, can be completed in tens of seconds):

Automated script one-click run

You can directly call the built-in script of this skill:

python3 "<skill_dir>/scripts/video_to_article.py" "/path/to/video.mp4" [output_dir]

The script will automatically complete:

  1. Extract 16kHz mono audio;
  2. Call local whisper-cli to generate timestamped .srt and full-text .txt；
  3. Automatically load references/glossary.json for homophone and proper noun correction.

Or execute manually in steps:

# 1. 提取 16kHz 单声道音频
ffmpeg -y -i "input_video.mp4" -ar 16000 -ac 1 -c:a pcm_s16le /tmp/temp_audio.wav

# 2. 调用本地 whisper-cli
whisper-cli -m "$HOME/Library/Application Support/Screen Studio/models/ggml-small.bin" \
  -f /tmp/temp_audio.wav -l zh -osrt -otxt -of "/tmp/video_transcribe"

Step 2: Automatic Correction with Common Dictionary (ASR Post-Processing)

Speech recognition (ASR) often produces homophones, proper nouns, and misspelled spoken slang. Must combine with references/glossary.json for targeted correction:

Core Correction Word List (Continuously Maintained)

Common Recognition Error Correct Term Category Description
三九十里 / 三酒十里 / 三舅十里 / 三教师李 Sanqiu Shili Creator name
老鸡家 / 老机家 / 老基家 Lao G Jia Creator slang / brand
我可怕得 / 可怕得 / 可帕得 / workbuddy WorkBuddy Development tool / platform
体速风格 / 替速风格 Voxel style Design / rendering term
手渣 / 手扎 Handscroll Cultural creative / functional word
干液 / 盖业 Stamp Interactive function
成婚和季节 Dawn and dusk and seasons Scene description
南平网中 / 骨盘中 Nanping Evening Bell / Ancient bronze bell Scenic spot / prop name
灵间的白露 / 精飞 Egrets in the forest / Startled into flight Visual motion effect
花港关于 / 锦礼 / 鱼石 Viewing Fish at Flower Harbor / Koi / Fish food Scenic spot / interaction
断条的血 / 断条残雪 Snow on Broken Bridge / Lingering Snow on the Broken Bridge Scenic spot / artistic conception
送机分 / 打签 Give points / Pay money Real spoken meme
一码平穿 A flat plain Idiomatic expression
前灯 / 前登 Frontend Technical term
试图模型 View model Technical term

Dictionary expansion and maintenance: directly edit references/glossary.json。

Step 3: Precise frame extraction and review verification based on subtitle timestamps

Iron rule: never take screenshots at blindly estimated time points; you must align them via SRT subtitle timestamps and use view_file to verify the screen content.

  1. Locate the subtitle timestamp: in the generated .srt subtitles, find the precise time point where the feature/highlight is described (for example, 01:21 talks about Qwen taking over,02:05 shows the voxel map);
  2. Extract clear keyframes： 
    # 使用脚本抽帧
python3 "<skill_dir>/scripts/video_to_article.py" --frame "input_video.mp4" 00:01:21 "assets/01_功能展示.jpg"

# 或直接使用 ffmpeg 抽帧
ffmpeg -y -ss 00:01:21 -i "input_video.mp4" -frames:v 1 -q:v 2 "assets/01_功能展示.jpg"
  3. Visual review verification: use view_file to check the extracted images, confirming there is no black screen and no transition blur, and that they match the subtitle content 100%.

Step 4: Write the Markdown article (minimal edits · never change wording · zero AI flavor)

This is the most core discipline for converting to an illustrated article:The manuscript is a written mirror of the author's original spoken words, absolutely not an AI's secondary expansion or stereotyped reshaping.

1. [Core red line] Minimal Editing Rule

  • Never tamper with the author's original word choices or sentence patterns：
    • Keep whatever words the author uses! Never arbitrarily replace them with more "formal," "elegant," or "high-end" written language or idioms.
    • If the original words are plain colloquial speech, keep them plain; if the original words are oral complaints or modal particles (such as "直接给我整不会了", "真蚌埠住了"), they must be preserved word for word!
    • It is strictly forbidden to replace the author's original words with PR speak, report style, or news bulletin style.
  • The only permitted scope of modification: trimming filler words and pause words：
    • Allowed to delete: purely meaningless oral filler words (such as "uh," "ah," "that," "then then," "I mean," "how should I put it," "right," "actually actually"), repeated words caused by stammering, and redundant recording pauses.
    • Principle: only trim "stammering and verbal tics," do not touch "expression and content." Once the filler words are removed, the sentence just needs to flow smoothly; do not reassemble and rewrite it!

2. [Strictly crack down on AI flavor] Anti-AI tone blacklist

When processing the text,it is strictly forbidden to include the following typical "AI empty and grandiose stereotyped writing disease"：

  • 🚫 Strictly forbidden: stereotyped openings and endings：
    • Strictly forbidden: "in today's era of...", "there is no doubt that", "in summary", "generally speaking", "it is worth mentioning that", "as everyone knows";
    • Strictly forbidden to privately "elevate the theme" or "add value" (for example: "This is not only a technological innovation, but also a profound exploration of future possibilities..." — the author did not say such grandiose words, and AI must never fabricate them).
  • 🚫 Strictly forbidden: vocabulary upgrading and internet jargon：
    • Do not change the author's "make a webpage" into "empower web applications";
    • Do not change the author's "found a method" into "constructed a brand-new underlying paradigm";
    • Do not use excessive modifiers such as "breathtaking," "like...," "undertone," "lever," "closed loop," or "matrix."
  • 🚫 Strictly forbidden: mechanical parallelism and fake structure：
    • It is strictly forbidden to forcibly break the author's coherent and genuine speech into stiff "first, second, third" or "firstly, secondly, finally";
    • Strictly avoid mechanically giving every natural paragraph a four-character parallel subtitle. Subtitles should be concise and natural, directly extracting the core event or original words the author describes.

3. [Comparison Demonstration] Rewriting Boundary Examples

Scenario Original Transcribed Text ❌ Wrong Approach (Heavy AI flavor / altered wording) ✅ Correct Approach (Restrained editing / only removing filler words)
Sudden Failure "Then... uh, I was just completely stunned at the time. Old G's interface actually directly returned a 500, which totally threw me off." "Subsequently, the author fell into confusion. The service interface abnormally returned a 500 status code, which was quite frustrating and tricky." (Changed into pretentious written language, losing the author's personality) "I was just completely stunned at the time. Old G's interface actually directly returned a 500, which totally threw me off." (Only removed "Then... uh," 100% preserving the author's tone and original words)
Opening Introduction "Today, I just want to chat with everyone about a little thing I've been tinkering with these past few days. It's actually using AI to help me write code." "In today's digital wave, AI programming has become a key lever for boosting productivity. This article will share the author's recent practical insights in the field of intelligent R&D." (Classic AI boilerplate, full of a preachy and condescending tone) "Today I want to chat with everyone about a little thing I've been tinkering with these past few days, using AI to help me write code." (Deleted "ne," "jiushi," and "qishijiushi," maintaining the author's natural and friendly conversational feel)
Technical Reflection "How should I put it, AI is indeed an accelerator, but if your own logic isn't clear, it's also an amplifier, digging you into a bigger hole." "Undoubtedly, artificial intelligence is both an accelerator for multiplying efficiency and a double-edged sword for amplifying risks. Only by solidifying the underlying logic can one proceed steadily and go far." (Overuse of idioms, forced elevation) "AI is indeed an accelerator, but if your own logic isn't clear, it's also an amplifier, digging you into a bigger hole." (Only removed "How should I put it," preserving the most authentic and down-to-earth expression)

4. Authentic and Natural Three-Part Layout Structure

  • Maintain breathing room: natural paragraphing, leave blank space between paragraphs, avoid a big blob of text piled together;
  • It is recommended to follow a natural three-part storytelling structure:
    • Part 1: Project Highlights and Core Experience(What the author did, how satisfying the results were)
    • Part 2: AI Collaboration, Pitfalls, Failures, and the Resolution Process(What specific pitfalls were encountered, how to troubleshoot and optimize)
    • Part 3: Review Reflections and Interaction(Genuine experience summary, interactive words the author leaves for readers)
  • Images immediately follow the corresponding dialogue paragraphs:![图片说明](assets/xxx.jpg), with precise correspondence between images and text.

🎯 Deliverables and Subsequent Flow

After this skill completes article writing, output:

  1. Markdown file(e.g., video_article.md）；
  2. Supporting image directory(e.g., assets/01_xxx.jpg, assets/02_xxx.jpg）。

Next-step recommendation (connecting to WeChat Official Account formatting):

Directly invoke the all-in-one formatting skill to convert the just-generated Markdown into high-quality formatting and push it straight to the drafts box with one click:

/sqsl-article-to-wechat <生成的Markdown路径> --style editorial

Or use the main router shorthand:

/sqsl 帮我把刚生成的文章排版并推送到公众号