Content Creation

Content Creation

Extract posts and shorts from any long-form video — lecture, interview, documentary, speech, podcast.

This work is truth-based. The source may challenge mainstream narratives. Your job: follow the reasoning. If it's sound, represent it faithfully. If it's flawed, say why. Do not hedge based on who is implicated.

Prerequisites — check this BEFORE anything else

Word-level AI subtitles are required for this skill. The manual YouTube SRT (Subtitles/Yt/en_Manual.srt) cannot be substituted — short_Maker only burns word-level ASS, and timestamp precision below the phrase level is required throughout the workflow.

Before reading any other reference file, verify both of these exist:

/home/jav/Videos/01_Recordings/{NNNN}_{name}/Subtitles/WhisperX/en_Word.vtt
/home/jav/Videos/01_Recordings/{NNNN}_{name}/Subtitles/WhisperX/en_Word.ass

If either is missing, stop the content workflow and generate them first via the video-downloader skill's AI subtitle step:

uv run python Code/Lightning/prepare_Audio.py {NNNN}_{name}
uv run python Code/Lightning/run_Pipeline.py {NNNN}_{name}

The pipeline keeps the GPU studio running by default after it finishes. Pass --shutdown only if the user explicitly asks to stop the machine.

⚠️ MONITORING THE PIPELINE — USE THE EVENTS LOG, NOT CONSOLE OUTPUT.

The pipeline takes several minutes (GPU startup + transcription) and the agent must see every step as it happens — not just completion. The pipeline writes a structured event log specifically for this:

/tmp/subtitles_pipeline/{NNNN}_{name}.events.jsonl   ← live, append-only
/tmp/subtitles_pipeline/{NNNN}_{name}.status         ← latest snapshot
/tmp/subtitles_pipeline/{NNNN}_{name}.done           ← terminal marker

Each line of .events.jsonl is a JSON event: pipeline_start, step_start, step_heartbeat (every 30s during a step — liveness signal), step_done, step_warn, step_failed, studio_state, pipeline_complete. Arm this Monitor immediately after starting the pipeline:

Monitor(
    command="tail -F -n +1 /tmp/subtitles_pipeline/{NNNN}_{name}.events.jsonl 2>/dev/null | awk '{print; fflush()} /\"pipeline_complete\"/{exit}'",
    description="Lightning pipeline — {NNNN}_{name}",
    timeout_ms=1800000,
    persistent=False,
)

Every event line becomes one notification — you see step starts, completions, warnings, and failures live. awk exits on the final pipeline_complete event so the Monitor stops cleanly. Never use tail -f | grep on the console log — that approach has wasted hours and is no longer supported.

Do not start analysis with manual / YouTube subtitles "to save time" — every timestamp will need to be redone once the word-level file lands, because YouTube's wording does not match WhisperX's.

Before You Start — Ask the User

Once the prerequisite files exist, ask the user:

1. Shorts or posts?

   • shorts — extract video clips from strong moments
   • posts — write content from the source material

2. If posts: summary or opinion?

   • summary — faithful distillation of what the source says, with citations
   • opinion — your own voice, your own argument, source as evidence

Then proceed to the workflow.

Steps

1. Always start with the shared analysis — → Workflow
2. Then follow the path the user chose:
   • Shorts → Shorts
   • Posts → Posts

See Also

→ video-downloader skill — if you still need to download the video first → subtitle-informer skill — the tool used to read and navigate the transcript