yt-replier

# YouTube Comment Replier Two-stage pipeline for managing YouTube comments on your channel - mirrors `/tiktok-replier`, but YouTube has an official API so there is **no Playwright**. Both reading comments and posting replies go through the YouTube Data API v3. Fully self-contained: own OAuth token (`token.json`), own auth module (`auth_yt.py`), own `data/` dir. No dependency on other skills. ## Files - `auth_yt.py` — OAuth (`youtube.force-ssl` scope). Reuses `~/credentials.json` client, caches token in `token.json` next to it. Run `python3 auth_yt.py` to authorize, `--reauth` to redo. - `monitor_yt.py` — fetches recent uploads, finds unreplied top-level comments, splits new ones into `data/inbox_yt.json` (manual) + `data/drafts_queue_yt.json` (auto keyword CTAs). Cron-friendly. - `queue_draft_yt.py` — moves a drafted reply from `inbox_yt.json` → `drafts_queue_yt.json`. - `reply_yt.py` — posts each reply in a queue via `comments().insert`, tracks done in `data/posted_yt.json`. Dry-run by default; `--post` to publish. - `data/inbox_yt.json` — comments needing manual reply (full metadata: cid, author, text, video_id, video_url, etc). - `data/drafts_queue_yt.json` — replies ready to post. - `data/posted_yt.json` — comment IDs already replied to (prevents double-posting). - `data/last_seen_yt.json` — dedupe state for the monitor. - `data/drafts_yt.md` — human-readable running log. ## The hourly cron + manual draft workflow (PRIMARY pattern) A cron runs `monitor_yt.py` every hour at :07. It produces three artifacts in `data/`: - **`inbox_yt.json`** — comments needing manual reply. This is what you ask Claude to help draft. - **`drafts_queue_yt.json`** — auto-drafted Skool-link replies for short keyword CTAs (System / Plan / Skill / Email / Routine / Tools / VFX / schedule / video / workflow). Already ready to post. - **`drafts_yt.md`** — human-readable running log. ### When the user says "any new youtube comments?" / "check my yt inbox" 1. Read `data/inbox_yt.json` (manual-needed) AND `data/drafts_queue_yt.json` (auto-drafts pending). 2. Show the user the breakdown — count of each plus the actual text of inbox items. 3. He picks one or all to draft replies for. ### When the user asks "draft a reply for the @username one" 1. Read the inbox entry for that comment. 2. Compose a reply in your tone (casual, helpful, drives to skool.com/the-ai-agency when relevant). Use `/harut` for conversion-sensitive wording. No em dashes. 3. Show the user the draft, get approval. 4. When approved, run: ``` python3 ~/.claude/skills/yt-replier/queue_draft_yt.py --cid <cid> --reply "<text>" ``` This appends to `drafts_queue_yt.json` AND removes from `inbox_yt.json`. ### When the user says "post them" ``` python3 ~/.claude/skills/yt-replier/reply_yt.py --post ``` `--post` is required to actually publish. Without it the script does a dry-run (prints what it would post). 10s spacing between posts. ## Standard workflow (manual / first-time) ### 0. Authorize (one time) ``` python3 ~/.claude/skills/yt-replier/auth_yt.py ``` A browser opens, you approve the YouTube scope, token saves to `token.json`. (The token was migrated from the `/yt-upload` skill on creation, so this is usually already done.) ### 1. Fetch unreplied comments ``` python3 ~/.claude/skills/yt-replier/monitor_yt.py ``` Scans the 30 most recent uploads, up to 50 top-level comments each. A comment is "unreplied" when it's top-level, not authored by your channel, and you haven't replied in that thread. New ones get split into inbox (manual) vs auto-queue (keyword CTA). ### 2. Build reply queue Show the user the unreplied list, confirm wording, then either: - auto-drafts are already in `drafts_queue_yt.json`, or - draft a manual reply and move it with `queue_draft_yt.py` (see above). ### 3. Post the replies **Always dry-run first, then test on 1, then batch:** ``` python3 reply_yt.py # dry-run (shows everything, posts nothing) python3 reply_yt.py --post --limit 1 # smoke test - post a single reply python3 reply_yt.py --post # batch (10s pause between each) ``` To post the auto-drafted keyword CTAs from a specific queue: ``` python3 reply_yt.py --post --queue ~/.claude/skills/yt-replier/data/drafts_queue_yt.json ``` ## Editing the auto-reply behavior `monitor_yt.py` holds the auto-draft logic near the top: - `SKOOL_LINK_REPLY` — the canned reply text for keyword CTAs. - `KEYWORD_REPLIES` — the keyword → reply map (System / Plan / Skill / Email / Routine / Tools / VFX / schedule / video / workflow). - `draft_reply_for()` — comments longer than 60 chars always go to the manual inbox; short ones matching a keyword get auto-drafted. ## Safety rules - **Default to dry-run**, then `--limit 1` on first run of any new queue. - **Always show drafts to you** before bulk posting (matches `feedback_confirm_before_scheduling.md` — confirm before posting even if approved earlier in the session). - **`posted_yt.json` tracks done IDs** so re-runs after a crash skip what already worked. - **No em dashes** in any reply (matches `feedback_no_em_dashes.md`). - **No automatic re-fetch + re-post loops** — you initiate each batch. ## Quota YouTube Data API v3: 10,000 units/day. Per monitor run is ~30-60 units; each posted reply (`comments.insert`) is 50 units. Plenty of headroom for normal use. ## Cron ``` 7 * * * * /Library/Frameworks/Python.framework/Versions/3.12/bin/python3 \ /Users/tylerreed/.claude/skills/yt-replier/monitor_yt.py \ >> /Users/tylerreed/.claude/skills/yt-replier/data/monitor_yt.log 2>&1 ``` This runs at :07 each hour (TikTok monitor runs at :05). The monitor posts a macOS notification with the new-comment count. It never posts replies on its own — posting is always a manual `reply_yt.py --post`.