SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsclaude-youtube-video-workflow

claude-youtube-video-workflow

Design & creation

View repository

Updated 10-16 00:20

About this skill

claude-youtube-video-workflow: Claude Code edition YouTube production master control

One-stop overview (long-form → short-form → dual-format info)

This Skill completes three major stages in one go, with 3 STOP checkpointsin between, and each checkpoint muststop and wait for the user to make the call; do not rush ahead automatically：

階段 1 長片：剪口播 → 轉字幕 → 🛑STOP1 疑慮字幕確認 → 時間碼硬關卡 → 標題 → 封面 → 長片 metadata → 打包
階段 2 短片：讀完整字幕 → 亮點偵測 → 🛑STOP2 三候選確認 → 切片組片 → 短片 metadata → 打包
（標題候選為 🛑STOP3，已含在各階段內）
  • 🛑 STOP 1(Stage 1 step 3): list questionable terms and wait for the user to proofread; only proceed after applying them.
  • 🛑 STOP 2(Stage 2): 3 candidate short-form versions, wait for the user to choose a number.
  • 🛑 STOP 3(step 4 / short-form step): title candidates, wait for the user to choose.
  • Hard checkpoint：validate_srt.py not passed (inconsistent segment count / timecodes don't match) → abort delivery and report back, must not force output.
  • Prevent word breaks: subtitle sentence splitting relies on resegment.py cut-point priority order; short-form clips must always align to subtitle paragraph boundaries,must not cut in the middle of a sentence。

The short-form stage is not mandatory: if the user doesn't want short-form, stop at the end of Stage 1. Only proceed to Stage 2 if short-form is wanted.

Read first

  • HANDOFF.md
  • CLAUDE.md
  • assets/style/reference-thumbnails.png
  • assets/style/cover-style.md
  • assets/persona/三師爸人物形象照.png

Claude Code-specific rules

  • Cover uses skills/cover-image/draw.py or Claude global ~/.claude/skills/draw/draw.py。
  • Requires OpenAI API Key:OPENAI_API_KEY or ~/.openai.env。
  • Cover generation must include --edit assets/persona/三師爸人物形象照.png。
  • Every cover must reuse the person's base reference photo; the person must not be carried over from the previous cover or a derivative image.
  • Claude's signature color is orange:#FF8C42、#FFA500、#FF6B35。
  • Claude output folders always have [Claude] suffix added.

Workflow

  1. Intake

    • Find raw/ new videos.
    • Create working/<video-id>/。

  2. Cut talking-head

    • Read skills/smart-cut/SKILL.md。
    • Run skills/smart-cut/scripts/smart_cut.py。
    • this speaker's tested sweet spot：--threshold 0.06 --margin "0sec,0.1sec"(cut about 80%, crisp pacing, no silence left at the start).
    • Starting suggestion: first run once with the tested sweet spot, show the user cut.mp4 to confirm the pacing; too aggressive → threshold 0.05, margin "0.2sec,0.2sec"; too loose → threshold 0.07.
    • Output working/<video-id>/<video-id>.cut.mp4。
    • After running, report "original length X:XX → new length X:XX (cut X%)".

  3. Transcribe subtitles

    • You can feed cut.mp4 directly to transcribe_groq.py(internally uses ffmpeg automatically), no need to extract audio first. Files < 25MB are taken directly by Groq; > 25MB will be automatically downsampled.
    • Read skills/audio-to-srt/SKILL.md。
    • Workflow: transcribe_groq.py → resegment.py → apply_vocab.py → Claude personally cleans the text segment by segment(read vocab.srt, fix typos per the rules, add punctuation, don't touch timecodes, keep segment count unchanged) → 🛑 STOP 1 questionable term confirmation(see below) → validate_srt.py [hard checkpoint] → srt_to_txt.py.
    • First create _subtitles/ subfolder, otherwise resegment.py will FileNotFoundError。
    • 🛑 STOP 1: Confirmation of questionable terms (no skipping allowed)
      1. Run python skills/video-editing-and-subtitles/scripts/find_dubious_terms.py "working/<video-id>/_subtitles/<video-id>.vocab.srt" --out "working/<video-id>/_subtitles/dubious_terms.md"to automatically extract paragraphs containing proper nouns (Claude/Gemini/Agent…) and common transliteration typos.
      2. Read dubious_terms.md, list the questionable paragraphs in the conversation, and stop to ask the user how to fix them. Do not guess and keep going.
      3. After the user replies, use python skills/video-editing-and-subtitles/scripts/finalize_subtitles.py "<vocab.srt>" --out "<clean.srt>" --replace "舊->新" --replace "段號:舊->新" to apply the corrections (for specific paragraphs, use only the 390:AGE->Agent format to avoid accidentally damaging things like the Gem in Gemini).
    • Timecode hard gate：validate_srt.py --raw <vocab.srt> --clean <clean.srt> If the segment counts are inconsistent or the timecodes do not match → abort, report the discrepancies, and do not deliver。
    • Common Whisper misrecognitions(remember to fix these when cleaning the text):
      • "燭光國中 / 竹光國中" → "光武國中"
      • "稅教材" → "數學教材"
      • "AI-Aging / AI 阿真 / AI 安進" → "AI Agent"
      • "租寫的範本" → "填寫的範本"
      • "微填" → "微調"
      • "半年段另存心" → "按年段另存新檔"
      • Cloud / Claude / Codex are easily confused (apply_vocab already handles most of this)
    • Output working/<video-id>/<video-id>.srt and .txt, and also copy out the .srt for convenient use in step 8.

  4. Generate titles and pause

    • After reading the cleaned text, .txt。
    • generate 10 title candidates to working/<video-id>/titles.md。
    • The title strategy should be grouped (to make it easy for the user to compare): pain point/liberation type, teaching/Know-how type, contrast/rhetorical question type, and concrete/case study type, with 2–3 of each.
    • Report the title list to the user,stop and wait for the user to choose a number. Do not pick one yourself and do not keep running.

  5. Create the Claude output folder

    • Clean Windows-illegal characters from the title:？！：／＼?!:/\\<>|"*。
    • Create output/<清洗後標題> [Claude]/。
    • Do not add [Claude]。

  6. Generate cover

    • Pre-cover SOP (all are required, do this every time)： a. Read assets/style/reference-thumbnails.png(look at the channel's existing 12 covers) b. Read assets/style/cover-style.md(read the full style guide) c. Read assets/persona/三師爸人物形象照.png(re-read the character reference photos every time; do not reuse old covers) d. Decide the main color based on the video's main subject: Claude=orange / Codex=blue / both=orange+blue
    • Claude version main color is fixed:#FF8C42、#FFA500、#FF6B35
    • Prompt template(already tested and effective; fine-tune according to the video content): 
      YouTube 封面 16:9，三師爸頻道科技教學風格。本支影片主角為 {AI_NAME}，採用{COLOR}主題配色。

背景：深海軍藍 (#0A1628) 漸層底，左下與右上有 {THEME_LIGHT_COLOR} 放射光線與粒子特效，細微網格紋理，整體科技未來感、強烈呼喚點擊感。

人物：畫面右側放置這位男性教師（戴眼鏡、黑色連帽外套），姿勢為{POSE}，臉部表情自信微笑，看向鏡頭。{COLOR} rim light 打在邊緣輪廓。人物樣貌、五官、髮型、外套、眼鏡全部延續輸入的人物基準照，不要更動。

主視覺（左側中下方）：{TOPIC_VISUAL}。電腦周圍有發光暈。

文字（左半邊）：
- 集數標籤（左上）：『{SERIES_LABEL}』白色細字
- 主標題（左中央）：『{TITLE_LINE_1}\\n{TITLE_LINE_2}』兩行排列，亮黃色 (#FFD700) 超粗黑體中文字，粗黑色描邊 5px，佔畫面寬度約 45%
- 副標題（主標下方）：『{SUBTITLE}』白色細字，搭配兩側裝飾橫線

整體：高對比、教學頻道氣口、適合教師受眾。
不要：純白底、自然風景、寫實生活感、3D 浮雕字。
    • Standard call: 
      python skills\cover-image\draw.py "<prompt>" --edit "assets\persona\三師爸人物形象照.png" --size 1536x1024 --quality low --name cover --outdir "output\<標題> [Claude]"
    • Run the cleanup: delete the timestamped filename version, keep cover.png one copy; use the Read tool to take a look at the cover and confirm the character's appearance is correct.

  7. Write metadata (according to marketing specifications)

    • First Read references/marketing-spec.md(the merged version of the txt v1.5 main specification + channel supplementary fields), and always use it as the standard.
    • Takeafter cleaning the text .txt transcriptgenerate; the material must echo the actual subtitle content. For long videos, metadata.md must include at least:
      1. 10 title candidates(Curiosity/Value/Pain Point three styles interspersed)
      2. Video description ≤300 characters: Hook (first two lines eye-catching) + 3 key knowledge points in bullet list + CTA; the first paragraph echoes the emotional tone of the title type selected by the user.
      3. Social post（FB / IG / Threads）：Emoji strictly prohibited, each paragraph ≤3 lines, long video 150–200 characters, sincere teacher tone, end with an interactive question.
      4. 15–20 SEO keywords + "Paste all at once" integrated tag block(separated by half-width commas, so the user can paste directly into YouTube backend).
      5. Chapter timestamps(based on subtitle segments, advisory in nature).
      6. Pre-publish checklist(title/thumbnail/subtitles/first 100 characters of description contain main keyword/tags/playlist/social sync).
    • Detailed formatting and short-video override rules are entirely based on references/marketing-spec.md .

  8. Packaging and self-check

    • The output folder should contain 5 files(file names without the [Claude] suffix):
      • <標題>.mp4(copied from working/ and renamed)
      • <標題>.srt
      • <標題>.txt
      • cover.png
      • metadata.md
    • Self-check checklist(check item by item after running):
      • The folder name ends with the [Claude] suffix
      • The folder contains all 5 files
      • The video file name equals the YouTube title (without the suffix)
      • cover.png opens and the person's appearance continues from the reference photo (take a look with Read)
      • cover.png main color is orange (not blue)
      • metadata.md contains the "paste all at once" integrated tag block
      • HANDOFF.md has been updated with this video's status
    • PowerShell reminder: when the path contains [Claude] , PowerShell treats square brackets as wildcards; use -LiteralPath or add single quotes. Bash does not have this problem.

  9. Update handoff

    • Update HANDOFF.md: completed items, output location, cover status, pending review items, next steps.

Pitfalls

  • OPENAI_API_KEY does not exist or the account is not verified, the cover will fail. Check order: environment variables → ~/.openai.env. Bash loading:export $(cat ~/.openai.env | xargs)。
  • skills/cover-image/draw.py requires the openai Python package.
  • gpt-image-2 requires OpenAI Individual verification; if unverified it will return 403 Organization must be verified。
  • auto-editor not being in PATH does not matter,smart_cut.py will automatically fall back to python -m auto_editor。
  • Groq transcription uploads audio to a third-party service; if the content is sensitive, stop first and ask the user.
  • Groq 25MB limit:transcribe_groq.py built-in ffmpeg automatically downsamples to 16kHz mono 32kbps.
  • resegment.py will not automatically create the output folder, first mkdir _subtitles。
  • Chinese file names, spaces,[Claude] square-bracket paths: in PowerShell use -LiteralPath or single quotes; in Bash double quotes are enough.
  • Contradictory cover prompt (for example, saying the person wears blue but the reference photo is a black jacket) → gpt-image-2 will compromise; in the prompt describe only position and scene, do not re-describe facial features or clothing attributes.
  • Every cover must re-read the reference photo, you cannot use the previous cover as input; the model will amplify the error, and the more it generates, the less it resembles the original.

Stage 2: Short clip highlights (only run if a short clip is needed)

After the long video finishes (already has cut.mp4 + clean.srt), if the user wants a short clip,take over skills/short-video-workflow/SKILL.md the complete workflow. Key checkpoints:

  1. Mandatory full read of working/<video-id>/<video-id>.srt: before extracting highlights, you must first read the entire subtitle file and understand the timecodes; it is not allowed to just look at a segment and grab it.
  2. Highlight detection: score each segment 0–3 according to the checklist such as pain points/rhetorical questions/numeric promises/strong assertions/twist reveals.
  3. 🛑 STOP 2: Confirm three candidate versions
    • Create 3 versions (A pain-point type / B curiosity type / C promise type hook), each with a three-act structure, ≤120s, written into working/<video-id>/shorts-candidates.md。
    • list the timecodes and three-act scripts of the three versions for the user, stop and wait for the user to choose A/B/C(or the user provides their own timecodes). Do not choose on your own.
  4. Slice and assemble：clip_cut.py slices (align to subtitle segment boundaries, do not cut in the middle of a sentence) → confirm ≤120s → add_end_card.py ending title card → burn_subtitles.py burn subtitles → make_vertical.py produce a 9:16 vertical version.
  5. Short clip title (🛑 STOP 3): produce 3 shorter and more sensational candidates and wait for the user to choose.
  6. Short clip metadata: likewise Read references/marketing-spec.md, apply the "C. Short clip differences" overrides (3 titles, description ≤150 characters,#Shorts required, add at the end of tags Shorts,短影片）。

Specs: 16:9 edit, ≤120s, primary color Claude orange, output output/<短片標題> [Claude] (Short)/, cover/person reference photos use the same set. Each short clip must also produce a 9:16 vertical version(for YouTube Shorts publishing): after burning subtitles, run make_vertical.py(blurred background fill, content not cropped, 1080×1920), output filename <短片標題>_直式9x16.mp4, the short clip folder has 7 files in total.

Example: result parameters of this project's first video

So you know what parameters have been tested OK:

  • Video:你還在手動填課程計畫嗎 AI Agent 教師工作流實演
  • original length 14:00 → smart-cut (threshold 0.06, margin "0sec,0.1sec") → 1:42 (cut 88%)
  • Subtitles 37 segments, 530 characters, total duration 1:41
  • 10 title candidates, divided into four strategy groups, user chose #7
  • Cover: orange theme, left side Claude AI interface + red X paper document, right side person pointing left
  • Complete final product:output/你還在手動填課程計畫嗎 AI Agent 教師工作流實演 [Claude]/