youtube-advisor — build a Claude Code advisor from any YouTube channel

Two-action user contract

The user does exactly two things:

1. Drops a channel URL/handle and a one-sentence intent in chat.
2. Uses the resulting `/<channel-slug>-advisor` skill to ask questions about that channel.

You (the AI agent) do everything in between. Never expose --since, --max, --out, bash, or any CLI flag to the user. Never paste a shell command for the user to run. Compose every command silently, run it via Bash, and report results in plain language.

When this skill triggers

• A YouTube channel URL, @handle, or playlist URL appears in the message.
• Creator name appears in a "build me an advisor on …" / "make a skill from …" / "make a YC advisor" framing.
• The user says "обнови yc-advisor", "refresh that advisor", "check for new videos".

Workflow

Step 1 — parse intent silently

Extract from the user's message:

• Channel(s) — URL, @handle, playlist URL, or a named creator.
• If a creator is named without a URL, search via yt-dlp ytsearch5:"<name>" (or yt-dlp --get-id ytsearch3:"<name> channel"), then ASK the user "I found @navalr — Naval Ravikant's channel — use that one?" before continuing. Don't guess.
• Selection hints:
• "last year" → --since 2025-MM-DD (today minus 1 year).
• "long videos only" → --title-exclude "shorts|#shorts" (already default — no extra flag).
• "the Founder Stories series" → --title-include "Founder Stories".
• Pasted video URLs → write IDs to /tmp/ids.txt, pass --ids /tmp/ids.txt.
• Intent sentence — for the LLM SKILL.md drafter (--intent "...").
• Language preference — match the chat language by default; pass --answer-language only if the user explicitly says.

If the message is ambiguous, ask AT MOST 3 short questions in plain language. NEVER list flag names. NEVER show CLI syntax.

Step 2 — confirm in plain language

Print a one-line summary:

"I'll fetch ~187 videos from @ycombinator (2022 → today), output to ~/.claude/skills/yc-advisor. Roughly 20 minutes. Go?"

Only ask about Whisper time if the estimated duration is > 15 minutes for the Whisper fallback.

Step 3 — execute via Bash (HIDDEN from the user)

youtube-advisor bootstrap \
  --channel @ycombinator \
  --since 2022-01-01 \
  --out ~/.claude/skills/yc-advisor \
  --from-natural-language --intent "YC startup advice with focus on hiring" \
  --answer-language auto \
  --quote-style auto \
  --yes

Run it via the Bash tool. The user does NOT see this command.

Step 4 — stream progress sparsely

Reuse the scriba progress-streaming pattern:

• Emit a short status message on each stage transition (resolving channel, ingesting, indexing, embedding, drafting SKILL.md, running evals).
• Emit on ~25% audio-processed boundaries for Whisper transcription (NOT every 10%, NOT on a timer).
• NEVER use /loop, ScheduleWakeup, or any periodic wake-up — they cost prompt-cache misses ($).

#### Live status from a side terminal

While bootstrap is running, you can peek at progress without tailing the verbose stdout:

bash <repo>/scripts/status.sh <advisor-dir>
# or
youtube-advisor status <advisor-dir>

Both read <advisor>/.progress.json and print a one-line snapshot, e.g. 🎬 ingesting · 4/187 · 02:14 · ETA 18:45.

Use this when the user asks "how far along is it?" — read the progress JSON, don't tail bash output. The JSON updates atomically on every stage change and per-video ingest.

Step 4 — polish the draft (YOU are the LLM)

Whenever bootstrap ran without ANTHROPIC_API_KEY (the default Claude Code case) — or when you passed --no-llm — the CLI writes a stub SKILL.md plus a hand-off file at <advisor>/.pending-llm-draft.json. Do not echo the postrun guide yet. Instead:

1. Read <advisor>/.pending-llm-draft.json (it contains the schema, ~8 transcript excerpts, the corpus metadata, the user intent, and the output paths).
2. Draft a polished SKILL.md directly from those samples — anti-patterns must be specific to THIS channel's actual failure modes (not generic clichés), and example_queries must be real questions this corpus can answer well.
3. Re-render SKILL.md and AGENTS.md at the paths listed in output_paths. Use the same Jinja templates the stub came from (templates/advisor.SKILL.md.tmpl, advisor.AGENTS.md.tmpl) — only the description/purpose/anti_patterns/example_queries change.
4. (Optional) Draft evals/benchmark.json (10–20 channel-grounded Q/A items with expected_quote_video_ids and must_terms) so the user can later run a quality check.
5. Delete `.pending-llm-draft.json` — that marker file's presence means "AI hasn't polished yet". Removing it signals the advisor is ready.

Only then echo the postrun guide.

Step 5 — show the postrun guide

When bootstrap exits, it prints the "what's next" guide. Echo it to the user AS-IS (verbatim from the CLI output). Do NOT add commentary about what flags you used. The user should see only the guide and a brief "Done!" framing.

Update flow

When the user says "обнови yc-advisor", "refresh the advisor", "check for new videos", "что нового на канале":

youtube-advisor update --advisor ~/.claude/skills/yc-advisor --yes

Same UX: never show the user the command. Echo the postrun summary. If the user provides a hint ("обнови, но только за эту неделю"), translate it to --since <date>.

Worked examples (for YOU the agent — never shown to the user)

Each example shows: user message → flags YOU compose → what the user sees.

Example 1 — bare URL + intent

User: https://www.youtube.com/@ycombinator. I want a YC advisor focused on startup hiring and fundraising.

You parse:

• channel = @ycombinator
• intent = "YC startup advice with focus on hiring and fundraising"
• language = match chat language (RU/EN auto)

You run:

youtube-advisor bootstrap \
  --channel @ycombinator \
  --out ~/.claude/skills/yc-advisor \
  --from-natural-language \
  --intent "YC startup advice with focus on hiring and fundraising" \
  --answer-language auto \
  --yes

User sees: confirmation summary → progress updates → postrun guide. NEVER sees the flags.

Example 2 — date hint, long videos only

User: pick interviews longer than 1 hour from 2024 onwards on @lexfridman

You parse:

• channel = @lexfridman
• since = 2024-01-01
• title_exclude default already drops shorts; "longer than 1 hour" is a length filter not yet in v1 — skip it and accept all matching videos.

You run:

youtube-advisor bootstrap --channel @lexfridman --since 2024-01-01 \
  --out ~/.claude/skills/lex-advisor \
  --from-natural-language --intent "Long-form Lex Fridman interviews" --yes

Tell the user: "Note: I'll pull videos from 2024 onwards. The shorts filter is on by default. Length-based filtering isn't a v1 feature, so very short interviews will also be included — let me know if you'd like to drop them manually later."

Example 3 — explicit video IDs

User: Use these specific videos: https://youtu.be/abc, https://youtu.be/def, https://youtu.be/ghi

You:

1. Extract IDs: abc, def, ghi.
2. echo -e "abc\ndef\nghi" > /tmp/youtube-advisor-ids.txt (via Bash).
3. Run:

youtube-advisor bootstrap --channel "<infer from one of the videos via yt-dlp>" \
  --ids /tmp/youtube-advisor-ids.txt \
  --max 3 \
  --out ~/.claude/skills/<slug>-advisor \
  --from-natural-language --intent "<ask user>" --yes

Ask the user briefly: "These three videos come from <channel>. Use that as the channel handle? And: what should this advisor be good at? (One sentence.)"

Example 4 — named creator, no URL

User: Build me an advisor on Naval

You:

1. Run yt-dlp --flat-playlist -J "ytsearch3:Naval Ravikant channel" | jq -r '.entries[] | .channel' to find candidates.
2. Ask: "I found @navalr — Naval Ravikant's channel. Use that one?"
3. On y, run bootstrap with --channel @navalr.

Example 5 — update with hint

User: обнови yc-advisor — также проверь нет ли новых видео за эту неделю

You parse:

• advisor = ~/.claude/skills/yc-advisor
• since override = 7 days ago.

You run:

youtube-advisor update --advisor ~/.claude/skills/yc-advisor --since 2026-05-29 --yes

Tell user the result via the postrun summary.

Rules

• Never show flags to the user. Always build the command silently. Run it via Bash. Echo the result.
• Never paste a `bash` command for the user to run. You run everything; you echo only natural-language summaries and the post-run guide.
• Match the user's language in conversation (RU / EN / etc.) — but the CLI itself is English-only.
• If Whisper time is estimated > 15 min, ask before continuing ("Some videos lack captions and would need ~3h of local transcription. Skip them, or wait?").
• Don't poll status on a timer. Reuse the scriba pattern: emit on stage change + 25% audio progress.
• End with the bootstrap_guide output unmodified. No additional commentary.
• Never use `/loop` or `ScheduleWakeup` for progress monitoring — they cost prompt-cache misses.
• Don't fabricate channel handles. If the user names a creator without a URL, always disambiguate via search first.

Installation prerequisites

If youtube-advisor is not installed (which youtube-advisor fails), do the install yourself — the user shouldn't have to leave chat. Clone into ~/.claude/skills/youtube-advisor (or ~/tools/youtube-advisor outside Claude Code), then run bash scripts/install.sh via the Bash tool. If ffmpeg is missing, install.sh will surface that — run brew install ffmpeg (macOS) or the platform equivalent and retry. Only ping the user if a step needs a sudo password or a non-trivial decision.

ANTHROPIC_API_KEY is not required when this skill is invoked from inside Claude Code — the agent in chat (you) drafts SKILL.md directly via the .pending-llm-draft.json hand-off (see Step 4). If the env var IS set, the CLI auto-uses the Anthropic SDK to pre-draft SKILL.md and the eval benchmark — that is purely an optimization, never a prerequisite. Pass --no-llm to force the hand-off path explicitly.

When YouTube blocks the fetch

If you see HTTP 429 / "Sign in to confirm you're not a bot" / _extract_player_response failed errors, YT's anti-bot is rate-limiting this IP. yt-dlp invocations already pass --remote-components ejs:github (the EJS JS-challenge solver) and --extractor-args youtube:formats=missing_pot (PO-token tolerance) by default — but the EJS solver needs `deno` on PATH at runtime. If which deno fails, install via brew install deno (macOS) or curl -fsSL https://deno.land/install.sh | sh (Linux) and retry. If you're still blocked after that, two more fixes:

Option 1 — Use your logged-in browser cookies (recommended): Pass --cookies-from-browser chrome (or firefox, safari, brave, edge). yt-dlp reads cookies straight from the browser profile — no copy-paste. With cookies set, the captions pool stays at concurrency 8.

Option 2 — Wait it out: The defaults already slow requests (1–5s random sleep) and retry on 429 with exponential backoff. Without cookies the captions pool also drops to concurrency 2 + a 0.5–2s jitter per call. A retry after a few minutes usually clears.

You can also set YOUTUBE_ADVISOR_COOKIES_BROWSER=chrome once in your shell rc, or drop a Netscape cookies.txt at ~/.config/youtube-advisor/cookies.txt, to default that browser/file for all advisors.