Agency Docs Updater

Agency Docs Updater — End-to-End Pipeline

When this skill is invoked, execute ALL steps below automatically in sequence. Do not stop to ask for confirmation between steps — run the full pipeline. Only pause if a step fails and cannot be recovered.

Step 1: Find Fathom Transcript

DATE=$(date +%Y%m%d)

Look for ~/Brains/brain/${DATE}-claude-code-lab-02.md. If it exists, read it and extract share_url and fathom_id from its YAML frontmatter.

If the file does NOT exist: 1. Run ~/.claude/skills/calendar-sync/sync.sh to sync today's calendar 2. Re-check for the file 3. If still missing, stop and report the issue

Store these variables for later steps:

• FATHOM_FILE = full path to transcript (e.g. ~/Brains/brain/20260207-claude-code-lab-02.md)
• SHARE_URL = the share_url from frontmatter
• MEETING_TITLE = the title from frontmatter (e.g. "Claude Code Lab 03 — Meeting 01: Знакомство и введение")
• DATE = YYYYMMDD string
• VIDEO_NAME = ${DATE}-claude-code-lab-02
• LAB_NUMBER = lab number from filename (e.g. 03)
• MEETING_NUMBER = meeting number from title or determined in Step 5 (e.g. 01)

Step 2: Download Video from Fathom

First check if ~/Brains/brain/${VIDEO_NAME}.mp4 already exists and is > 1MB. If so, skip this step.

Otherwise:

cd ~/Brains/brain && python3 ~/.claude/skills/fathom/scripts/download_video.py \
  "${SHARE_URL}" --output-name "${VIDEO_NAME}"

After download, verify the mp4 exists and is > 1MB:

ls -la ~/Brains/brain/${VIDEO_NAME}.mp4

If download fails, stop and report. This is a long-running operation (may take several minutes for a 2h video).

Step 3: Upload to YouTube via videopublish

Prerequisites check (run before first invocation):

cd ~/ai_projects/youtube-uploader && node -e "require('playwright')" 2>/dev/null || (npm install playwright && npx playwright install chromium)

cd ~/ai_projects/youtube-uploader && \
python3 process_video.py \
  --video ~/Brains/brain/${VIDEO_NAME}.mp4 \
  --fathom-transcript ${FATHOM_FILE} \
  --title "${MEETING_TITLE}" \
  --upload

Key notes:

• Always pass `--title` with the title from the Fathom transcript frontmatter. Without it, the LLM generates a generic/wrong title (e.g. "Клод Код Лаб" instead of the proper meeting name)
• Do NOT use `source venv/bin/activate` — youtube-uploader has no venv, run python3 directly
• --fathom-transcript makes it skip video transcription (uses Fathom transcript instead)
• --upload triggers YouTube + Yandex.Disk upload
• Handles: metadata generation, thumbnail creation, YouTube upload, Yandex upload
• Thumbnail generation requires playwright npm package and Chromium browser

Extract YouTube URL from stdout — look for the line:

✓ YouTube video: https://www.youtube.com/watch?v=VIDEO_ID

If not found in stdout, check the metadata JSON:

cat ~/ai_projects/youtube-uploader/processed/metadata/${VIDEO_NAME}.json

Store YOUTUBE_URL for the next steps.

This step is long-running (10-30 minutes depending on upload speed). Run in background with run_in_background: true.

If the upload fails or stalls mid-way, resume from the upload step only (skips metadata/thumbnail regeneration):

python3 process_video.py --video ~/Brains/brain/${VIDEO_NAME}.mp4 \
  --fathom-transcript ${FATHOM_FILE} --title "${MEETING_TITLE}" --upload --resume-from upload

Parallelization: Start Step 4 (summary generation) while Step 3 upload runs in background. The summary does not depend on the YouTube URL.

Step 4: Generate Fact-Checked Russian Summary

Read the full Fathom transcript from ${FATHOM_FILE}.

Generate a comprehensive Russian-language summary of the meeting with these requirements:

• Structured with ## section headers
• Bullet points for key concepts
• Code examples where relevant (keep code/paths in English)
• All technical terms in English (MCP, Skills, Claude Code, YOLO, vibe coding, etc.)
• Comprehensive enough to serve as meeting notes
• Exclude personal scheduling details (e.g. "X will be on vacation next week") — only include content relevant to the meeting topic itself

Then use the Task tool with claude-code-guide subagent to fact-check all Claude Code feature claims in the summary:

Launch claude-code-guide agent to fact-check this summary about Claude Code features.

Verify:
- Subagent types and their capabilities
- Tool names and parameters
- Feature availability and limitations
- Best practices mentioned

Correct any inaccuracies.

After fact-checking, save the corrected summary to the scratchpad directory as summary.md.

Step 4b: Update YouTube Video Description

Do this after both Step 3 (upload) and Step 4 (summary) are complete.

The description generated by process_video.py's built-in LLM (Groq) is generic and low-quality for Russian content. Instead, generate the YouTube description yourself using the summary from Step 4.

Determine the meeting page URL: https://agency-lab.glebkalinin.com/docs/claude-code-internal-XX/meetings/NN (where XX = lab number, NN = meeting number).

Generate a YouTube description in this format:

${MEETING_TITLE}

[1-2 sentence overview of the meeting in Russian]

В этом видео:
- [bullet point 1 — key topic covered]
- [bullet point 2]
- ...
- [bullet point 8-10 max]

Материалы и конспект занятия:
https://agency-lab.glebkalinin.com/docs/claude-code-internal-XX/meetings/NN

Сообщество AGENCY — практики AI-агентов:
https://agency-lab.glebkalinin.com

#ClaudeCode #AI #Anthropic #Claude #AIагенты #программирование

Then update the video on YouTube using the API via auth.py (which manages token.pickle with youtube.force-ssl scope):

cd ~/ai_projects/youtube-uploader && PYTHONPATH=. python3 -c "
from auth import get_authenticated_service
youtube = get_authenticated_service()
resp = youtube.videos().list(part='snippet', id='VIDEO_ID').execute()
snippet = resp['items'][0]['snippet']
snippet['title'] = MEETING_TITLE
snippet['description'] = DESCRIPTION
snippet['tags'] = ['Claude Code', 'Claude', 'Anthropic', 'AI', 'AI агенты', 'программирование', 'Claude Code Lab', 'MCP']
youtube.videos().update(part='snippet', body={'id': 'VIDEO_ID', 'snippet': snippet}).execute()
"

Key notes:

• auth.py uses youtube.force-ssl scope which covers both uploads and metadata updates
• token.pickle persists across sessions — no browser auth needed after initial setup
• If token expires, auth.py auto-refreshes it
• Extract VIDEO_ID from YOUTUBE_URL
• Generate the bullet points from the summary content, not from the transcript directly

Add video to playlist — after updating metadata, add the video to the lab's playlist:

cd ~/ai_projects/youtube-uploader && PYTHONPATH=. python3 -c "
from auth import get_authenticated_service
youtube = get_authenticated_service()

# Find playlist matching 'Claude Code Lab XX'
resp = youtube.playlists().list(part='snippet', mine=True, maxResults=50).execute()
playlist_id = None
for p in resp['items']:
    if p['snippet']['title'] == 'Claude Code Lab LAB_NUMBER':
        playlist_id = p['id']
        break

if playlist_id:
    youtube.playlistItems().insert(part='snippet', body={
        'snippet': {
            'playlistId': playlist_id,
            'resourceId': {'kind': 'youtube#video', 'videoId': 'VIDEO_ID'}
        }
    }).execute()
    print(f'Added to playlist: {playlist_id}')
else:
    print('Playlist not found — create it manually on YouTube first')
"

Known playlist IDs (for reference):

• Claude Code Lab 03: PLZNP0SKU2Sqic4njcSQemmyrVLHZ4C_iT
• Claude Code Lab 02: PLZNP0SKU2SqizoM9_7jMjgae2cvjGN9bQ
• Claude Code Lab: PLZNP0SKU2Sqga_EjxAW_OxnqaUk1ZGfXU

Replace LAB_NUMBER with the two-digit lab number (e.g. 03). The playlist name must match exactly (e.g. "Claude Code Lab 03").

Step 5: Run update_meeting_doc.py

python3 ~/.claude/skills/agency-docs-updater/scripts/update_meeting_doc.py \
  ${FATHOM_FILE} \
  "${YOUTUBE_URL}" \
  ${SCRATCHPAD}/summary.md

The script auto-detects:

• Lab number from filename (claude-code-lab-XX -> XX)
• Target docs dir: ~/Sites/agency-docs/content/docs/claude-code-internal-XX/
• Next meeting number from existing files in meetings/
• Presentation files from ~/ai_projects/claude-code-lab/presentations/lab-XX/
• Summary language (auto-translates to Russian if needed)

IMPORTANT: Meeting number detection — The script picks the next available number, but placeholder MDX files may already exist for future meetings with dates pre-filled. Before using the script's auto-detected number: 1. List existing MDX files in meetings/ 2. Check if any placeholder file already has today's date in its content (e.g. grep -l "14 февраля" meetings/*.mdx) 3. If a placeholder exists for today's date, use --update flag or -n NN to target that file instead of creating a new one

Output: MDX file at ~/Sites/agency-docs/content/docs/claude-code-internal-XX/meetings/NN.mdx

After the script runs, post-process the generated MDX:

1. Strip appended presentation content: The script appends Marp presentation markdown from presentations/lab-XX/ which contains HTML comments (<!-- _class: lead -->) that break MDX compilation. Remove everything after the summary section (after the last --- separator following the summary). The MDX should only contain: frontmatter, video section, and summary.

2. Copy lesson HTML to public: If ~/ai_projects/claude-code-lab/lesson-generator/${DATE}.html exists, copy it to ~/Sites/agency-docs/public/${DATE}-claude-code-lab-XX.html (where XX is the lab number). Then add a link in the MDX video section:

   **Материалы:** [Презентация занятия](/${DATE}-claude-code-lab-XX.html)

3. Replace frontmatter placeholders:

• [Название встречи] -> actual meeting title derived from transcript content
• [Краткое описание встречи] -> brief description
• [Дата встречи] -> formatted date from the transcript

4. Verify build locally before committing:

   cd ~/Sites/agency-docs && npm run build 2>&1 | tail -5

If build fails, fix the MDX (common issues: HTML comments, unescaped < or { characters) and retry.

Step 6: Commit and Push

cd ~/Sites/agency-docs && git add . && git commit -m "Add meeting NN" && git push

Replace NN with the actual meeting number from Step 5 output.

This triggers Vercel deployment automatically.

Step 7: Verify Vercel Deployment

Wait ~90 seconds after push, then check deployment status:

gh api repos/glebis/agency-docs/commits/COMMIT_HASH/status --jq '{state, total_count}'
gh api repos/glebis/agency-docs/commits/COMMIT_HASH/statuses --jq '.[0] | {state, description}'

• If state: success — deployment is live.
• If state: failure — check the build error locally with cd ~/Sites/agency-docs && npm run build 2>&1 | tail -20, fix, and re-push.
• If state: pending — wait another 60 seconds and re-check.

Pipeline Summary

After completion, report: 1. Fathom transcript path 2. Downloaded video path 3. YouTube URL 4. Generated MDX path 5. Git commit hash 6. Vercel deployment status (success/failure)

Error Recovery

• Step 1 fail (no transcript): Run calendar-sync, retry once. If still missing, stop.
• Step 2 fail (download): Check if share_url is valid. Report ffmpeg errors.
• Step 3 fail (upload): Check YouTube OAuth tokens, Playwright installation. Report specific error. Do NOT try source venv/bin/activate.
• Step 4 fail (summary): Proceed with English summary if Russian generation fails.
• Step 5 fail (MDX): Check that fathom_url exists in frontmatter. Check docs_dir exists.
• Step 6 fail (git): Report conflict details. Do not force-push.

Reference: CLI Interfaces

download_video.py

python3 download_video.py <share_url> --output-name <name_without_ext>

Downloads to current directory as <name>.mp4.

process_video.py

python3 process_video.py --video <path> --fathom-transcript <path> --upload

Flags: --upload (YouTube+Yandex), --yandex-only, --skip-thumbnail, --language <code>, --title <override>.

update_meeting_doc.py

python3 update_meeting_doc.py <fathom_file> <youtube_url> <summary_file> [-n NN] [-l ru|en|auto] [--update]

Learnings

2026-02-07

Context: First full pipeline run for meeting 08

What Worked:

• --resume-from upload avoids re-generating metadata/thumbnails on upload retry
• Parallelizing summary (Step 4) with YouTube upload (Step 3) saves 10+ minutes
• gh api repos/OWNER/REPO/commits/HASH/status reliably checks Vercel deploy status
• Local npm run build catches MDX errors before pushing

What Failed:

• YouTube upload speed degraded from 5.4 MB/s to 0.15 MB/s mid-upload (3 attempts needed)
• update_meeting_doc.py appended lab-01 Marp presentation with <!-- --> comments — broke MDX build
• Pushed without local build check — caused Vercel deploy failure, required fix commit

Known Issues:

• presentations/lab-XX/ may contain presentations from wrong meetings — always strip appended content
• YouTube API upload speed is highly variable and not controllable from client side
• MDX does not support HTML comments (<!-- -->), unescaped <, or bare { characters

2026-02-14

Context: Pipeline run for meeting 10 (Agent SDK workshop)

What Worked:

• Parallelizing summary with YouTube upload continues to save significant time
• Date-matching existing placeholder MDX files correctly identifies target meeting number
• Local npm run build caught issues before push

What Failed: 1. venv activation — source venv/bin/activate fails because youtube-uploader has no venv. Fixed: run python3 directly 2. Playwright missing — ERR_MODULE_NOT_FOUND: Cannot find package 'playwright' during thumbnail generation. Fixed: added prerequisite check in Step 3 3. Wrong meeting number — update_meeting_doc.py auto-detected meeting 13 (next available), but meeting 10 was a placeholder for today's date. Fixed: added date-matching guidance in Step 5 4. Personal scheduling details in summary — "X will be on vacation" included in published summary. Fixed: added exclusion rule in Step 4 5. Appended Marp content — still happens despite being documented. Reinforced in Step 5 post-processing

Key Fix: Always check existing MDX files by date content before trusting auto-detected meeting number. Use -n NN flag to override.

2026-03-04

Context: Pipeline run for Lab 03, Meeting 01

What Worked:

• Full pipeline completed successfully (transcript, download, upload, summary, MDX, deploy)
• Fathom recorded as "Impromptu Zoom Meeting" — found by date, renamed correctly

What Failed: 1. Wrong YouTube title — Without --title flag, LLM generated "Клод Код Лаб" instead of proper meeting name. Fixed: added --title "${MEETING_TITLE}" to Step 3 using frontmatter title 2. YouTube OAuth scope too narrow — youtube.upload scope can't update video metadata (title/description) after upload. Needed youtube.force-ssl scope for videos().update() call 3. Appended Marp content — Still happens (3rd time). Truncation via Edit tool only replaced first match, leaving 1100+ lines of Marp. Fixed: rewrote entire file with Write tool

Key Fix: Always pass --title from Fathom frontmatter title field to process_video.py. The LLM metadata generator produces poor/generic titles for Russian-language content.