bilingual-video-note-enhancer

Files (8)

• SKILL.md
• bilingual-translation-style.md · references/
• obsidian-note-shape.md · references/
• screenshot-selection-rules.md · references/
• build_contact_sheet.sh · scripts/
• export_selected_frame.sh · scripts/
• extract_video_segments.sh · scripts/
• update_frame_map.py · scripts/

SKILL.md (first lines)

# Bilingual Video Note Enhancer

Enhance an English video note into a bilingual Obsidian note that stays readable, compact, and visually useful. Keep the original English source, add a concise Chinese TL;DR near the top, insert Chinese translations directly under the corresponding English blocks, and replace weak screenshots with readable content frames.

Default to Obsidian-compatible Markdown and stable asset handling. Preserve the note's existing structure unless the user explicitly asks for a larger rewrite.

## Preserve Note Identity

Do not invent a new free-form title for the note unless the user explicitly asks for retitling.

Title rules:
- Preserve the existing note title by default.
- Do not append generic suffixes such as `- transcript note`, `- bilingual note`, or similar machineish labels.
- If the source title is obviously a temporary placeholder and the user wants cleanup, normalize it to the simplest stable topic title rather than a descriptive sentence.

## Workflow

1. Read the target note.
2. Detect frontmatter, section headings, timestamp guide, transcript layout, and image embeds.
3. Decide whether the note is:
- a video note with timestamps and screenshots
- a transcript note with timestamps but no screenshots
- an English article-like note that still needs bilingual enhancement

## Choose Enhancement Mode

Pick the lightest mode that fully satisfies the task. Do not automatically enter the screenshot workflow just because the note contains timestamps.

- **Translation-only mode**: Use when the user wants Chinese TL;DR plus paragraph-level translation, and the note either has no screenshots or does not need visual upgrades.
- **Screenshot-refresh mode**: Use only when the note already contains screenshot embeds, the screenshots are weak, or the user explicitly asks for better visual inserts.
- **Article-note mode**: Use when the note is article-like or transcript-lite and should gain bilingual enhancement without being forced into a video-note structure.
- **Fallback mode**: Use when screenshot work would normally be relevant but local video tooling, source clips, or parallel review support are unavailable. In that case, still complete the bilingual note enhancement and preserve existing embeds rather than inventing fake screenshot updates.

Mode rules:
- If the task can be completed well with TL;DR + translation only, stop there.
- Do not invent screenshot work for article-like notes.
- Do not convert an existing note into a brand-new layout unless the user explicitly asks for a larger rewrite.

## Minimum Inputs

Choose the lightest path that the available materials can actually support.

- **Translation-only mode** and **Article-note mode** require only the target note path or note content. Existing screenshots, video URLs, or local clips are optional and should not be requested just because timestamps are present.
- **Screenshot-refresh mode** requires the target note plus at least one reusable visual source that covers the needed timestamps: the original video URL, a local video file path, or an existing local clip or asset directory.
- If the note has timestamps but no accessible video URL, local video file, or reusable clip assets, do not start the screenshot workflow. Stay in Translation-only mode or Fallback mode and preserve the existing embeds.
- If the user explicitly wants screenshot upgrades but cannot provide a usable visual source, say the screenshot path is blocked and still complete the bilingual enhancement without fabricating new frames.

## Add Chinese TL;DR

Add a short Chinese summary near the top of the note after the title.

Default output for YouTube or video notes:
- Add only the Chinese TL;DR.
- Do not add extra `abstract`, `info`, `note`, or `Key Takeaways` blocks unless the user explicitly asks for them or the note already depends on that structure.
- Prefer 3 to 5 concise bullets or one short paragraph, whichever matches the note better.

TL;DR requirements:
- Summarize the whole piece, not only the opening.
- Prefer concrete takeaways over vague praise.
- Keep terminology consistent with the body translation.

Read [references/obsidian-note-shape.md](references/obsidian-note-shape.md) if you need note placement rules.

## Insert Chinese Translations

Insert each Chinese translation directly below the English source block it translates.

Translation rules:
- Do not prefix the Chinese block with `译文：`.
- Preserve timestamps, speaker cues, and section headings.
- Keep the English source intact unless the user asked for cleanup.
- Translate faithfully, but compress obvious filler words when the spoken English is repetitive.
- Keep product, model, and tool names in their conventional forms.

If the note already uses a pattern like:

```md
**12:30** · English source paragraph
```

then produce:

```md
**12:30** · English source paragraph

对应中文译文段落
```

Read [references/bilingual-translation-style.md](references/bilingual-translation-style.md) for style rules and consistency checks.

## Upgrade Screenshots

Use this section only when the note has timestamp screenshots or the user explicitly wants visual inserts.

Before starting screenshot work, check these preconditions:
- The note actually contains screenshot embeds or clearly needs visual inserts.
- A reusable visual source is available: `yt-dlp` against the source URL works, a local video file exists, or an existing local clip or asset directory can be reused for the target timestamps.
- There is enough support to review candidate frames well. Prefer parallel subagents when available, but do not block the whole task if they are unavailable.

Default screenshot workflow:
1. Build a timestamp list from the note.
2. Use `yt-dlp` to download a short high-resolution local clip for each timestamp, or download the full video first when the source is short and local re-seeking is simpler.
3. Generate a contact sheet from the local clip.
4. Prefer parallel subagents to review candidate contact sheets and identify the best content frames.
5. If subagents are unavailable, review the contact sheets in the main thread and explicitly treat that as a fallback path with lower confidence.
6. Review the returned picks locally and confirm the frame that best represents each section topic.
7. Export the final image from that same local clip.
8. Replace the existing image file in place or create a stable new asset path.
9. Write or update a frame map JSON in `output/<topic>-<date>/`.

Preferred scripts:
- `scripts/extract_video_segments.sh`
- `scripts/build_contact_sheet.sh`
- `scripts/export_selected_frame.sh`
- `scripts/update_frame_map.py`

Selection policy:
- Prefer slide, diagram, workflow, code, title page, or clearly legible UI over speaker-only frames.
- For YouTube talks, lectures, and recorded presentations, projected PPT or handout content is usually the primary target. Treat speaker-only shots as low-value unless the speaker's physical action is itself the thing being explained.

From AgenticAIPlan/AgenticAISkills, MIT.

More from AgenticAIPlan/AgenticAISkills