Video Review

/video-review

When to use

The user gives you a batch of video links (Instagram, Facebook, YouTube), each paired with a free-text instruction, e.g.:

https://instagram.com/reel/abc123 - verify the claim they make about X, screenshot at 25s if true
https://youtube.com/watch?v=xyz - just give me a digest, no ticket
https://facebook.com/watch/def456 - if they're wrong about Y, make a JOB ticket

Links + instructions are pasted directly into the skill call. There is no queue file, every invocation is self-contained.

Pipeline

1. Parse the batch

Split the pasted input into (link, instruction) pairs. If a link has no instruction, treat it as "digest only, no ticket" (the safe default).

2. Route each link

Read the instruction text yourself and classify it. This decision stays on Claude, it's a cheap read of a short string, not worth delegating.

• Needs real transcription if the instruction asks to verify a specific claim precisely, quotes or paraphrases something exact that was said, or asks for a screenshot/frame at a timestamp.
• Digest only otherwise, a general "what is this about", "summarize it", "is this worth my time" type instruction.

3. Check the Ollama Track A watcher (before any digest-only dispatch)

Run:

Get-ScheduledTask -TaskName JobHunter_Ollama_DropfolderWatcher | Select-Object TaskName, State

• If Running, proceed.
• If not, cycle it once:

  Stop-ScheduledTask -TaskName JobHunter_Ollama_DropfolderWatcher -ErrorAction SilentlyContinue
  Start-ScheduledTask -TaskName JobHunter_Ollama_DropfolderWatcher
  

  Re-check state. If still not Running, stop and tell the user the watcher is down and needs manual attention (registering a new task needs an interactive credential prompt only they can supply). Do not silently fall back to fetching every digest-only link yourself, that defeats the point of the delegation and burns tokens without saying so.

Never invoke run_ollama_batch.ps1 directly, self-invocation is blocked. Only the watcher process runs it.

4. Dispatch digest-only links to Track A

For each digest-only link, write a task file to D:\Projects\Agent_Orchestrator\ollama_dropfolder\tasks\<slug>.txt:

<WorkingDir, e.g. D:\Projects\Agent_Orchestrator>
glm-5.2:cloud
Fetch and summarize this video link for a content-review batch: <link>
Give: what the video claims/covers, tone, and anything that looks like a
factual claim worth double-checking. Plain ASCII only, no em dashes.

Poll D:\Projects\Agent_Orchestrator\ollama_dropfolder\done\<slug>_done.txt for the result (reasonable timeout, e.g. 3-5 minutes per link; these run sequentially through the watcher, a big batch takes longer, say so if it's running long).

5. Real transcription for links that need it

For each verification-needed link:

py -3 D:\Projects\Shared_Content\video_transcriber\transcribe.py <url> --model small --outdir <dir> --frames N

• Always py -3, never bare python/python3 (Job Hunter's venv lacks yt-dlp/faster-whisper; the system py -3 has them).
• Use --model small for non-English audio, tiny can return empty transcripts on Arabic.
• Before requesting a screenshot at a specific timestamp, check the video's actual duration first (transcribe.py reports it); a timestamp past the end will fail. If the instruction's requested timestamp is out of range, say so in the verdict rather than silently skipping it.
• Convert any /tmp/...-style output path to a real Windows path before reading it with the Read tool.

6. Verdict (Claude only, never Ollama)

For every link, compare what was fetched/transcribed against that link's own instruction and produce:

• Claim status: verified true / verified false / unverifiable (say why)
• What the instruction asked for: did you actually deliver it (screenshot taken, claim checked, etc.)
• One-line verdict summary

Sanity-check any tool/product name against known real names before calling a claim unverifiable. Whisper systematically mis-transcribes "Claude" as "Cloud" in tech-review content, if a transcript says "Cloud" in a context that's obviously about an AI assistant, treat it as "Claude" and verify accordingly.

7. Conditional Linear ticket

Only create a ticket when the link's own instruction explicitly asks for one ("make a ticket if...", "log this", "ticket it"). Never invent ticket- worthiness yourself.

Route by instruction content:

• Mentions Job Hunter / CV / application context → Job Hunter, JOB-:

  python -m src.linear_push --update TICKET-ID -d "..."
  

  run from D:\Projects\Job_hunter\.
• Mentions Azure / chocolates / business context → Azure, SOLO-:

  python linear_push.py --update TICKET-ID -d "..." --project "Azure Projects"
  

  run from D:\Projects\Azure\Azure_projects\.
• If the instruction says to create a new ticket rather than update one, follow the same script's create path, still --update-style append semantics apply once it exists. Never pass --replace.

8. Output

• 1-3 links total: give every verdict directly in the chat response, no file written.
• More than 3 links: build one consolidated report using templates/report_template.html (fill in one section per link: link, instruction, fetch method used, verdict, screenshot if any, ticket link if any), save to reports/<date>_batch_review.html in the current project directory (create reports/ if missing), then open it:

  start reports/<date>_batch_review.html
  

  This is a plain local file, not a hosted Artifact.

Known gotchas

• Bash's bare python resolves to the wrong venv, always use py -3 for video_transcriber calls.
• /tmp/... paths need conversion to real Windows paths before Read can load them.
• Whisper mis-transcribes "Claude" as "Cloud", sanity-check before flagging a claim unverifiable.
• The watcher check is mandatory before every digest-only dispatch batch, not just the first link, it can go down mid-session.