process-youtube-transcript

name	process-youtube-transcript
description	Process YouTube transcript notes in the Obsidian vault inbox — apply first-mention wikilinks, generate summaries, and enrich frontmatter. Use this skill whenever the user says "process youtube", "process transcripts", "enrich youtube notes", "summarize youtube", or wants to process clippings-tagged YouTube notes in the inbox.

Process YouTube Transcript

You are a transcript enrichment agent for an Obsidian vault. Your job is to find YouTube transcript notes in 0.inbox/, apply first-mention wikilinks, generate summaries, and insert them into the note. You do NOT file or move notes — that stays with the inbox-processor.

Re-Processing Guard

Notes track which agents have processed them via the ai-processed frontmatter property — a YAML list of agent names. Skip notes where ai-processed already contains process-youtube-transcript. Only process notes that:

1. Are in 0.inbox/
2. Have a clippings tag
3. Have a YouTube URL in the source: frontmatter field (matching youtube.com or youtu.be)
4. Do NOT have process-youtube-transcript in their ai-processed list

If all matching notes are already processed, report "No new YouTube transcripts to process" and stop.

The ai-processed Pattern

This is a vault-wide convention. Any agent that enriches a note appends its name to the ai-processed list in frontmatter:

ai-processed:
  - process-youtube-transcript
  - inbox-processor

This creates an audit trail of which agents have touched the note. To force re-processing by this skill, the user removes process-youtube-transcript from the list.

Step 0: Preflight

Verify Obsidian is reachable:

obsidian vault

If it fails, report the error and stop.

Step 1: Find Unprocessed Notes

Search for candidates:

obsidian search query="clippings" path="0.inbox/" limit=20

For each result, read its frontmatter to check:

• source: contains youtube.com or youtu.be
• ai-processed does NOT contain process-youtube-transcript

If no unprocessed notes found, report and stop.

Show the user a summary table of what will be processed:

Note	Author	Published
Full AI Trading Bot Using Claude Code!	Alex Carter	2026-04-04

Step 2: Build Vault Vocabulary

Run the shared vocabulary script — a single obsidian eval call that returns all entities as JSON in ~0.2s:

bash .claude/skills/shared/build-vocab.sh

Returns { people: [...], projects: [...], areas: [...], tags: [{tag, count}] }. This replaces the old 4-call pattern (obsidian search × 3 + obsidian tags counts).

Focus on proper nouns and domain terms. Cache this vocabulary for the entire batch run — pass it to each subagent.

Step 3: Process Each Transcript

For each unprocessed note, delegate to a subagent (Agent tool with model: "sonnet"). Each transcript can be 10K-35K+ words, so subagents protect the main context. Sonnet handles transcript summarization and entity linking well, and keeps token costs manageable for large batches.

Subagent Instructions

The subagent receives:

• The note path
• The vault vocabulary list
• Instructions for linking and summarization

The subagent must:

3a. Read the Full Note

Use Read tool to get the complete note content. Note the structure:

---
frontmatter
---
![](youtube-embed)

## Transcript
...content...

3b. Apply First-Mention Wikilinks

Kepano's "link on first mentions" strategy — create a [[wikilink]] the first time each entity appears in the transcript body. All subsequent mentions stay as plain text.

Rules:

• Cross-reference entities against the vault vocabulary
• If an entity matches an existing vault note, use the exact note name: [[Kirk Henderson]]
• If the entity doesn't have a vault note but is a meaningful proper noun (company, product, technology, person), still link it: [[New Entity]]
• Use aliases when the transcript uses a short form: [[Claude Code|Claude]]
• Match case-insensitively but preserve original casing in display text
• Don't link inside timestamp markers (e.g., **0:00** ·)
• Don't link common/generic terms — only proper nouns, products, companies, technologies, people
• Don't over-link: if uncertain, leave as plain text
• The author field in frontmatter already has wikilinks from the clipper — don't duplicate those in the body unless the author is mentioned by name in the transcript text

3c. Generate Summary

Produce a summary section with:

1. Overview (3-5 sentences): What is this video about? What's the creator's angle/thesis?
2. Key Takeaways: Bulleted list of the most important points, tools, or techniques covered
3. Tools & Technologies Mentioned: Bulleted list of specific products, platforms, APIs, etc.
4. Timestamps of Interest: 3-5 notable moments with timestamps and brief descriptions (for quick navigation)

3d. Insert Summary into Note

Use the Edit tool to insert the summary between the YouTube embed and ## Transcript.

Find this pattern in the note:

![](https://www.youtube.com/watch?v=...)

## Transcript

Replace with:

![](https://www.youtube.com/watch?v=...)

# Summary

## Overview
...

## Key Takeaways
...

## Tools & Technologies
...

## Timestamps of Interest
...

---

## Transcript

Also apply the wikilink changes to the transcript body using Edit tool.

3e. Update Frontmatter

Mark as processed by appending to the ai-processed list:

obsidian property:set name="ai-processed" value="[process-youtube-transcript]" path="0.inbox/Note Name.md"

If ai-processed already exists with other entries, append rather than replace. Read the current value first and construct the updated list.

Also update:

obsidian property:set name="fileClass" value="researchNote" path="0.inbox/Note Name.md"

Add source/youtube tag if not already present. Keep the existing clippings tag — the inbox-processor will handle it during filing.

Subagent Error Handling

If a subagent fails on a transcript:

• Log the error
• Do NOT add process-youtube-transcript to ai-processed — the note remains eligible for re-processing
• Continue with the next transcript
• Report the failure in the final summary

Step 4: Report

After all transcripts are processed, show:

Note	Entities Linked	Status
Full AI Trading Bot Using Claude Code!	8	Processed
Claude Code + TradingView	5	Processed

Total: X notes processed, Y entities linked Linked entities: flat list of all wikilinked terms across the batch Errors: any failures

Run History

After a successful batch, append a JSON entry to zz_config/skill_history/process-youtube-transcript.md via obsidian append (create with obsidian create ... template="Base Template" on first run).

Resolve hostname first, in its own Bash call, so it doesn't appear as $(...) in the content= arg (shell substitution triggers a permission prompt even when obsidian append * is allowlisted):

hostname -s

Then substitute the literal value into the JSON:

{"date": "YYYY-MM-DD", "host": "<literal-hostname>", "count": N, "notes": ["0.inbox/...", ...], "total_words": N_across_all_transcripts}

Skip the append when no unprocessed transcripts were found.

Edge Cases

• Very long transcripts (500+ lines): The subagent handles this — that's why each transcript gets its own subagent
• No transcript section: If ## Transcript heading is missing, skip the note and report it
• Already has a Summary section: If # Summary already exists but ai-processed doesn't list this skill (shouldn't happen normally), add the skill to ai-processed and skip
• Duplicate entity names: If "Claude" could mean Claude (AI) or Claude (person), prefer the most contextually relevant one
• Author already wikilinked in frontmatter: The clipper adds [[Author Name]] in the author: field. Don't re-link in the body unless the name actually appears in the transcript text
• Notes without clippings tag: Ignore — this skill only targets clippings-tagged YouTube notes
• Re-processing: Remove process-youtube-transcript from the ai-processed list to allow re-processing