Yt2bb

yt2bb — YouTube to Bilibili Video Repurposing

Overview

Six-step pipeline: download → transcribe → translate → merge → burn subtitles → generate publish info. Produces a video with hardcoded bilingual (EN/ZH) subtitles and a publish_info.md with Bilibili upload metadata.

When to Use

• User provides a YouTube URL (single video or playlist) and wants a Bilibili-ready version
• User needs bilingual EN-ZH subtitles burned into video
• User wants to repurpose English video content for Chinese audience

Quick Reference

Step	Tool	Command	Output
0. Update	git	Auto-check for skill updates	—
1. Download	yt-dlp	yt-dlp --cookies-from-browser chrome -f ... -o ...	{slug}.mp4
2. Transcribe	whisper*	srt_utils.py check-whisper then transcribe	{slug}_{lang}.srt
2.5 Validate	srt_utils.py	srt_utils.py validate / fix	{slug}_{lang}.srt (fixed)
3. Translate	AI	SRT-aware batch translation	{slug}_zh.srt
4. Merge	srt_utils.py	srt_utils.py merge ...	{slug}_bilingual.srt
4.5 Style	srt_utils.py	srt_utils.py to_ass --preset netflix|clean|glow	{slug}_bilingual.ass
5. Burn	ffmpeg	ffmpeg -c:v libx264 -vf ass=...	{slug}_bilingual.mp4
6. Publish	AI	Analyze content, generate metadata	publish_info.md

Update check

Throttle to one check per 24 hours per installation; never mutate the skill directory without explicit user consent. SKILL_DIR resolved here is reused by later pipeline steps for script paths.

1. If <this-skill-dir>/.last_update exists and is less than 24 hours old, skip this step entirely.

2. Otherwise, fetch the latest tag from upstream:

   git -C <this-skill-dir> ls-remote --tags origin 'v*' 2>/dev/null \
     | awk '{print $2}' | sed 's|refs/tags/||' \
     | sort -V | tail -1
   

3. Compare with this skill's metadata.version from the frontmatter. If the upstream tag is strictly newer (semver), tell the user one line and ask:

   "A newer version of this skill is available: vX.Y.Z → vA.B.C. Want me to git pull?"

   If they say yes, run git -C <this-skill-dir> pull --ff-only. Refresh .last_update either way so the prompt doesn't repeat for 24 hours.

4. If upstream is the same or older, refresh .last_update silently and continue.

5. On any failure (offline, not a git checkout — e.g. ClawHub-installed copy, read-only path, no permission), swallow the error silently and continue with the user's task. Do not mention the failure.

Resolve SKILL_DIR for use by later pipeline steps:

# Find skill directory (works across Claude Code, OpenClaw, Hermes, Pi)
SKILL_DIR="$(find ~/.claude/skills ~/.openclaw/skills ~/.hermes/skills ~/.pi/agent/skills ~/.agents/skills ~/myagents/myskills -maxdepth 2 -name 'yt2bb' -type d 2>/dev/null | head -1)"

Pipeline Details

Step 1: Download

Single video:

slug="video-name"  # or: slug=$(python3 "$SKILL_DIR/srt_utils.py" slugify "Video Title")
mkdir -p "${slug}"
yt-dlp --cookies-from-browser chrome \
  -f "bv*[ext=mp4]+ba[ext=m4a]/b[ext=mp4]" \
  -o "${slug}/${slug}.mp4" "https://www.youtube.com/watch?v=VIDEO_ID"

Playlist / series:

yt-dlp --cookies-from-browser chrome \
  -f "bv*[ext=mp4]+ba[ext=m4a]/b[ext=mp4]" \
  -o "%(playlist_index)03d-%(title)s/%(playlist_index)03d-%(title)s.mp4" \
  "https://www.youtube.com/playlist?list=PLAYLIST_ID"

After downloading, rename each folder to a clean slug and run Steps 2–6 for each video sequentially.

• -f "bv*[ext=mp4]+ba[ext=m4a]/b[ext=mp4]": ensure mp4 output, avoid webm
• %(playlist_index)03d: zero-padded index to preserve playlist order
• If --cookies-from-browser fails, export cookies first — see Troubleshooting

Step 2: Transcribe

First run the environment check to detect your platform and get a tailored whisper command:

python3 "$SKILL_DIR/srt_utils.py" check-whisper

This auto-detects OS, GPU (CUDA/Metal/CPU), memory, and installed backends, then recommends the best backend + model for your hardware. If memory detection is unavailable, it falls back conservatively instead of assuming a low-memory machine. Use the command it prints.

Manual fallback (openai-whisper, works everywhere):

src_lang="en"      # Change to ja/ko/es/etc. based on source video
whisper_model="medium"  # check-whisper recommends the best model for your hardware
whisper "${slug}/${slug}.mp4" \
  --model "$whisper_model" \
  --language "$src_lang" \
  --word_timestamps True \
  --condition_on_previous_text False \
  --output_format srt \
  --max_line_width 40 --max_line_count 1 \
  --output_dir "${slug}"
mv "${slug}/${slug}.srt" "${slug}/${slug}_${src_lang}.srt"

Supported backends:

Backend	Best for	Install
mlx-whisper	macOS Apple Silicon (fastest)	pip install mlx-whisper
whisper-ctranslate2	Windows/Linux CUDA, or CPU (~4x faster)	pip install whisper-ctranslate2
openai-whisper	Universal fallback	pip install openai-whisper

Model selection (auto-recommended by check-whisper):

• tiny — fast draft, low accuracy, CPU-friendly (~1 GB)
• medium — default, good balance (~5 GB)
• large-v3 — best accuracy, recommended for JA/KO/ZH source (~10 GB)

Notes:

• --language: explicitly set to avoid misdetection; supports en, ja, ko, es, etc.
• --word_timestamps True: more precise subtitle timing
• --condition_on_previous_text False: prevent hallucination loops
• If output is garbled or repeated, add anti-hallucination flags — see Troubleshooting

Step 2.5: Validate & Fix (optional)

python3 "$SKILL_DIR/srt_utils.py" validate "${slug}/${slug}_${src_lang}.srt"
# If issues found:
python3 "$SKILL_DIR/srt_utils.py" fix "${slug}/${slug}_${src_lang}.srt" "${slug}/${slug}_${src_lang}.srt"

Step 3: Translate

Read {slug}_{src_lang}.srt and translate to Chinese. Critical rules:

These rules are modeled on the Netflix Simplified Chinese Timed Text Style Guide; follow them to produce broadcast-grade subtitles.

1. Keep SRT format intact — preserve index numbers, timestamps (--> lines) exactly as-is
2. 1:1 entry mapping — every source entry must produce exactly one translated entry (same count)
3. Optimize for bottom subtitles — keep each Chinese entry to 1 line whenever possible so the final bilingual subtitle stays compact near the bottom of the frame
4. Max 16 full-width characters per line (Netflix SC spec). Prefer 12–16; if a cue is very short (< 1 s) compress further so reading speed stays ≤ 9 characters/second
5. Shorten with judgment, not mechanically — remove filler words, repeated subjects, weak interjections, and redundant politeness before dropping key meaning
6. Match subtitle duration — the line must feel readable within the time on screen; if the cue is very short, compress more aggressively
7. No trailing punctuation on Chinese cues — drop ending 。, ！, ？; keep mid-sentence ，, 、, ； only when they add clarity
8. Use full-width Chinese punctuation inside cues (，。！？、；：); use 「」 for inner quotes, not "" or ''
9. Half-width digits and Latin — numbers, units, product names, and code identifiers stay half-width (GPT-4, 30fps, 2026); only punctuation is full-width
10. Line-break discipline — never break after function words (的, 了, 吗, 呢, 吧, 啊); never split an English phrasal unit across a line break; keep modifiers with their heads
11. Keep terminology consistent — technical terms, names, product names, and recurring phrases should be translated the same way across batches. Maintain an inline glossary if needed
12. Adapt, don't transliterate — preserve register, tone, and intent over literal word matching; idioms become natural Chinese equivalents
13. Translate in batches of 10 entries — output each batch in valid SRT format, then continue
14. Do NOT merge or split entries — maintain original segmentation
15. Save as {slug}/{slug}_zh.srt

Step 4: Merge

python3 "$SKILL_DIR/srt_utils.py" merge \
  "${slug}/${slug}_${src_lang}.srt" "${slug}/${slug}_zh.srt" "${slug}/${slug}_bilingual.srt"

Step 4.25: Netflix Lint (recommended)

Run lint on the merged bilingual SRT to catch Netflix Timed Text Style Guide violations that validate doesn't cover — reading speed (CPS), per-line length, inter-cue gaps, and line count.

python3 "$SKILL_DIR/srt_utils.py" lint "${slug}/${slug}_bilingual.srt"

Defaults (all overridable via flags):

Rule	Threshold	Flag
Reading speed (English)	≤ 17 CPS	--max-cps-en
Reading speed (Simplified Chinese)	≤ 9 CPS	--max-cps-zh
Min cue duration	833 ms (5/6 s)	--min-duration-ms
Max cue duration	7000 ms	--max-duration-ms
Min inter-cue gap	83 ms (2 frames @ 24 fps)	--min-gap-ms
Max chars/line (English)	42	--max-chars-en
Max chars/line (Chinese, full-width)	16	--max-chars-zh
Max lines per cue	2	—

Severity model:

• Errors (exit code 2): duration out of bounds, CPS over limit, > 2 lines per cue. These break Netflix acceptance and should be fixed before burning.
• Warnings (exit 0 unless errors also exist): per-line length, tight gaps. These are recommendations — address if feasible, but they don't block delivery.

When CPS errors fire, the fix is almost always upstream — go back to Step 3 and rewrite the offending Chinese entry to fit the time window. Do not solve CPS by extending the cue past the source's spoken duration.

Agent-friendly output:

python3 "$SKILL_DIR/srt_utils.py" lint "${slug}/${slug}_bilingual.srt" --format json

Returns {ok, error_count, warning_count, issues: [{index, code, severity, message}, ...]} for programmatic filtering.

Step 4.5: Style — Convert to ASS

Convert the bilingual SRT to an ASS file. ASS enables per-line color, font size, and glow effects that are impossible with SRT force_style. Layout rule: subtitles always stay at the bottom. Default stack: ZH on the upper line of the bottom stack, EN on the lower line. The presets are tuned to keep the block readable while reducing overlap risk with lower-screen content.

IMPORTANT — Ask before proceeding. Present the preset table below to the user and ask which style they prefer. Do NOT silently pick a default. If the user has no preference, use clean.

Available presets:

Preset	Look	Best for
netflix	Pure white text, thin black outline, soft drop shadow, no box — modeled on the Netflix Timed Text Style Guide	Professional, broadcast-grade look. Best default for documentaries, interviews, long-form content, and anything that should feel "streaming-platform native". Use with --font "Source Han Sans SC" on Linux / "PingFang SC" on macOS for closest Netflix Sans feel
clean	Yellow text on gray box — golden ZH + light yellow EN, semi-transparent light gray background	Readability safety net for busy or mixed-brightness footage where netflix's outline-only text could get visually lost. The gray box guarantees a readable contrast pad
glow	Yellow ZH + white EN with colored glow — bright yellow ZH + white EN, blurred outer glow, no background box	Entertainment, vlogs, energetic edits. Most eye-catching, but weakest on bright or busy backgrounds

Example prompt to user:

字幕有三套样式可选：

1. netflix — 纯白字体 + 细黑描边 + 柔和阴影（默认推荐，Netflix 专业观感，适合纪录片/访谈/长内容）
2. clean — 黄色字体 + 灰色半透明底框（亮背景或花背景的兜底选项，底框保证对比度）
3. glow — 黄色/白色字体 + 彩色外发光（更抢眼，适合娱乐/Vlog）
4. 自定义 — 提供 .ass 样式文件，完全控制字体、颜色、大小（可用 Aegisub 可视化编辑）

选哪个？默认推荐 netflix；如果画面特别花哨或底部信息多，可改用 clean。

# Netflix-grade default (white + outline + soft shadow), ZH on top
python3 "$SKILL_DIR/srt_utils.py" to_ass \
  "${slug}/${slug}_bilingual.srt" "${slug}/${slug}_bilingual.ass" \
  --preset netflix

# Gray-box fallback for busy backgrounds, EN on top
python3 "$SKILL_DIR/srt_utils.py" to_ass \
  "${slug}/${slug}_bilingual.srt" "${slug}/${slug}_bilingual.ass" \
  --preset clean --top en

# Vibrant glow (B站 entertainment style)
python3 "$SKILL_DIR/srt_utils.py" to_ass \
  "${slug}/${slug}_bilingual.srt" "${slug}/${slug}_bilingual.ass" \
  --preset glow

Custom style file — for full control, provide an external .ass file with your own [V4+ Styles] section. It must contain styles named EN and ZH, or to_ass will fail early with a validation error. You can design styles visually with Aegisub and export.

python3 "$SKILL_DIR/srt_utils.py" to_ass \
  "${slug}/${slug}_bilingual.srt" "${slug}/${slug}_bilingual.ass" \
  --style-file my_styles.ass

Optionally add ; en_tag= and ; zh_tag={\blur5} comment lines in the .ass file to inject ASS override tags per language.

Font by platform (pass with --font, ignored when using --style-file):

Platform	Flag
macOS	--font "PingFang SC" (default)
Linux	--font "Noto Sans CJK SC"
Windows	--font "Microsoft YaHei"

Other options:

• --top zh|en — which language on the upper line of the bottom stack (default: zh)
• --res WxH — video resolution (default: 1920x1080)

Readability notes for all presets:

• Presets stay bottom-aligned at all times; they do not move to the top automatically
• Font size, outline, and vertical margins scale with --res so 720p and 1080p keep similar visual balance
• clean is the safest choice when you must keep subtitles at the bottom in every shot

Step 5: Burn Subtitles

Use the ass= filter (not subtitles=) — all styling comes from the ASS file.

ffmpeg -i "${slug}/${slug}.mp4" \
  -vf "ass='${slug}/${slug}_bilingual.ass'" \
  -c:v libx264 -crf 23 -preset medium \
  -c:a copy "${slug}/${slug}_bilingual.mp4"

• -c:v libx264 -crf 23: good quality with reasonable file size
• -preset medium: balance between speed and compression (use fast for quicker encode)
• No force_style needed — styles are embedded in the ASS file

Step 6: Generate Publish Info

Based on the video content (from {slug}_{src_lang}.srt and {slug}_zh.srt), generate {slug}/publish_info.md.

All output in this file must be in Chinese (targeting Bilibili audience).

# Publish Info

## Source
{YouTube URL}

## Titles (5 variants)
1. {Suspense/question style — spark curiosity}
2. {Data/achievement driven — emphasize results}
3. {Controversial/opinion style — spark discussion}
4. {Tutorial/practical style — emphasize utility}
5. {Emotional/relatable style — connect with audience}

## Tags
{~10 comma-separated keywords covering topic, technology, domain}

## Description
{3-5 sentences summarizing core content and highlights}

## Chapter Timestamps
00:00 {chapter name}
...

Generation rules:

• Title style must match Bilibili conventions: conversational tone, suspense hooks, liberal use of symbols (【】, ?, !)
• Tags should cover both Chinese and English keywords for discoverability
• Timestamps extracted from {slug}_bilingual.srt at topic transition points
• Description needs a strong hook — first two sentences determine whether users expand to read

Output Structure

{slug}/
├── {slug}.mp4              # Source video
├── {slug}_{src_lang}.srt   # Source language subtitles
├── {slug}_zh.srt           # Chinese subtitles
├── {slug}_bilingual.srt    # Merged bilingual
├── {slug}_bilingual.mp4    # Final output
└── publish_info.md         # Bilibili upload metadata

Utility: srt_utils.py

python3 "$SKILL_DIR/srt_utils.py" merge en.srt zh.srt output.srt          # Merge bilingual
python3 "$SKILL_DIR/srt_utils.py" merge --dry-run en.srt zh.srt output.srt # Pre-check without writing
python3 "$SKILL_DIR/srt_utils.py" validate input.srt                       # Check timing issues
python3 "$SKILL_DIR/srt_utils.py" fix input.srt output.srt                 # Fix timing/overlaps (multi-pass)
python3 "$SKILL_DIR/srt_utils.py" slugify "Video Title"                    # Generate slug
python3 "$SKILL_DIR/srt_utils.py" to_ass input.srt output.ass              # Convert to styled ASS (default: clean, ZH on top)
python3 "$SKILL_DIR/srt_utils.py" to_ass --dry-run input.srt output.ass    # Pre-check without writing
python3 "$SKILL_DIR/srt_utils.py" to_ass input.srt output.ass --preset glow --top en
python3 "$SKILL_DIR/srt_utils.py" to_ass input.srt output.ass --style-file custom.ass  # User-defined styles
python3 "$SKILL_DIR/srt_utils.py" check-whisper                    # Detect platform, recommend whisper backend + model

Common Mistakes

• Mismatched entry counts: Merge fails by default — fix translation or use --pad-missing to pad
• Font not found: Ensure PingFang SC is installed (macOS default) or substitute (see Troubleshooting)

Troubleshooting

yt-dlp: Cookie Auth Failure

--cookies-from-browser chrome requires Chrome to be closed (or uses a snapshot of the profile). If it fails:

# Export cookies once, then reuse the file
yt-dlp --cookies-from-browser chrome --cookies cookies.txt --skip-download "URL"
yt-dlp --cookies cookies.txt -f "bv*[ext=mp4]+ba[ext=m4a]/b[ext=mp4]" -o "${slug}/${slug}.mp4" "URL"

For 429 / rate-limit errors, add --sleep-interval 3 --max-sleep-interval 8.

whisper: Wrong Language or Hallucination Loops

Symptoms: repeated phrases, garbled characters, or near-empty SRT despite clear audio.

whisper "${slug}/${slug}.mp4" \
  --model medium \
  --language "$src_lang" \
  --condition_on_previous_text False \
  --no_speech_threshold 0.6 \
  --logprob_threshold -1.0 \
  --compression_ratio_threshold 2.0 \
  --output_format srt \
  --output_dir "${slug}"

If language is still misdetected, the audio likely has long silence or non-speech segments — add --vad_filter True to suppress them.

ffmpeg: Font Not Found / CJK Boxes

Pass the correct font via --font in the to_ass step (Step 4.5). The ASS file embeds the font name, so ffmpeg needs it installed at burn time.

Platform	Font	Install
macOS	PingFang SC	pre-installed
Linux	Noto Sans CJK SC	sudo apt install fonts-noto-cjk
Linux (alt)	WenQuanYi Micro Hei	sudo apt install fonts-wqy-microhei
Windows	Microsoft YaHei	pre-installed

Regenerate the ASS file with the correct --font flag, then re-run the burn step.

Privacy & Data Flow

• Browser cookies: Step 1 uses yt-dlp --cookies-from-browser chrome to access age-gated or private videos. This reads Chrome cookies locally — no cookies are transmitted beyond YouTube's own servers. To avoid this, export cookies to a file first (see Troubleshooting above).
• Transcripts & translation: Step 3 (translate) and Step 6 (publish info) are performed by the AI agent in the conversation. Transcripts are sent to whatever model/service the agent uses (e.g. Claude API). If the video contains sensitive content, use a local model for those steps.
• Auto-update check: The pre-flight step runs git fetch to check for skill updates. It does not auto-pull or execute remote code.
• No telemetry: srt_utils.py makes no network requests. All processing (SRT parsing, merging, ASS generation, hardware detection) is fully local.