# Analyze Stream Recording > Internal helper for youtube-stream-description. Transcribes a local stream recording on-device (ffmpeg + mlx_whisper) and returns a recap summary, a YouTube-valid chapter list, and any links mentioned aloud. Not for direct use; invoked by youtube-stream-description when the user supplies a video file path. - Skill: `ryanwelcher/analyze-stream-recording` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add ryanwelcher/analyze-stream-recording` - Raw SKILL.md: https://api.skillmd.com/api/skills/ryanwelcher/analyze-stream-recording/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: ryanwelcher (https://skillmd.com/u/ryanwelcher) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/ryanwelcher/analyze-stream-recording --- # Analyze Stream Recording You are running in a forked context. The parent conversation will see **only your final message**, so everything you read here stays here. Your job is to turn a long transcript into a few hundred tokens the parent can draft from. Input: `$ARGUMENTS` is the path to a local video/audio file (mp4/mov/mkv/wav). ## Steps 1. **Transcribe.** Run this in the **foreground** with the Bash tool's `timeout` set to `600000`: ```bash bash ${CLAUDE_SKILL_DIR}/scripts/transcribe-file.sh "$ARGUMENTS" ``` - **Never** use `run_in_background`, Monitor, or anything else that waits for a later notification. This fork ends the moment you stop calling tools, and nothing can wake it again. A result that arrives after that is lost, and the parent gets nothing. Do not end your turn until you have the three sections below or an `ERROR:` line. - It prints only pointers: `CONDENSED: `, `LINES`, `WORDS`, `SRT: `. It never prints the transcript. - If it prints `PENDING: ...`, the transcription is still running in a detached job. Run the **exact same command** again right away, in the foreground. It picks up the same job and does not restart it. Repeat until you get pointers or an `ERROR:` line. - If it prints an `ERROR:` line (missing `ffmpeg` or `mlx_whisper`, bad path, failed job), return that line verbatim as your entire response and stop. Do **not** attempt to install anything; the parent handles that. - Results are cached per file, so a repeat run on the same recording returns in seconds. Apart from `PENDING`, never re-run it to "check" something. - For a very long stream (over ~3 hours) pass a second argument of `300` for 5-minute buckets. - The decoder is biased toward WordPress vocabulary. For an off-topic stream, set `WHISPER_PROMPT` to a comma-separated list of the names and terms that matter before running. 2. **Read the condensed transcript** at the `CONDENSED` path. One `[HH:MM:SS] text` line per bucket. If `LINES` is large, read it in chunks with `offset`/`limit`; do not skip the tail, since the wrap-up is where the "what got solved" lives. 3. **Derive chapters.** Buckets are raw material, not the chapter list. Merge adjacent buckets covering the same work. YouTube silently ignores the whole list if any rule fails: - First chapter starts at `00:00`. - At least three chapters. - Each at least 10 seconds long, listed in ascending order. - One per line, `MM:SS Label`, or `HH:MM:SS Label` once past an hour. - Aim for 6–12 chapters on a 2-hour stream. - Never one chapter per bucket. Use at most one chapter per ~4 minutes of stream (a 44-minute stream gets 11 at most), and a chapter changes only when the work changes. If the stream returns to an earlier thread, fold that into the story instead of reopening it as its own chapter. - Labels describe what happens, not the clock: "Setting up theme.json" beats "Part 2". 4. **Return exactly the three sections below and nothing else.** No preamble, no transcript excerpts, no quotes longer than a sentence, and no file paths. ## Return format ``` ## Recap <2–4 past-tense sentences: what was built, what broke, what got solved. Raw material for a hook, not the hook itself.> ## Chapters 00:00