Fb Watch

/fb-watch — finishing the teasers for the owner's Facebook posts

> Why. The owner sometimes writes a post STRAIGHT onto his Facebook wall, bypassing the pipeline (voice → episode → tiers). > Such a post ends up WITHOUT a teaser in the RU chat / X (EN) → reach is lost. This watchdog catches > "there is an authored Facebook post with no teaser yet" and writes the teasers immediately. Draft-first until armed.

Tools (a thin orchestrator, AK-47)

• Reading the wall = the Claude-in-Chrome MCP (a live logged-in Facebook tab, locally on the hub). Reuses the pattern from /fb-reply (in-page javascript_tool, since Facebook hides names behind the MCP boundary).
• ⛔ Do NOT read the profile feed (consensus #0223a74a, 07-20): it is virtualized and renders ONE post — the neighbours stay eternal skeletons, scrollHeight is constant, and real scrolling does not help. One run hid 2 content posts out of 5 that way.
• ✅ The harvesting rail = the Content Library: https://www.facebook.com/professional_dashboard/content/content_library/ → a table of posts, FULL text (truncation is CSS-only), exact date-time, metrics. Parsing: document.body.innerText → lines, anchor on "Published", the text is the line before, the date the line after; accumulate across scrolls (dedupe by the first 60 characters).
• ⚠️ The Content Library does NOT give permalinks (the row menu only has Edit/Delete). To get them: on the wall, do a real computer hover over the post timestamp → Facebook fills pfbid into href (before hovering the href is empty; synthetic MouseEvents and React-fiber tricks do NOT work — Facebook uses a custom renderer key).
• UPGRADE (owner approved 2026-06-30, waiting on credentials): as soon as secrets\fb_graph.env exists (FBUSERTOKEN, user_posts) and python ~/.claude/scripts/fb_posts_poll.py check is green → read the wall through python ~/.claude/scripts/fb_posts_poll.py posts --limit 10 --out posts.json (Graph API, more reliable than the browser, no ban risk) INSTEAD of Claude-in-Chrome. See [[decision-social-posting-stack]] §6.
• Detector/ledger/cap = python ~/.claude/scripts/fb_teaser_watch.py (0 tokens, dedup + daily cap + kill switch). State: $IMPORTS_ROOT/content-factory/fb_teaser_ledger.json + fb_watch_config.json.
• Teasers are written by the top-tier model in-session (the owner's voice), palettes: _STYLE.md / _CRAFT.md + style-reality-show.md. Drafts → _imports\content-factory\fb-teasers\.md.
• RU posting = Telegram MCP send_message, the secondary work account (the creator), chat ** (the RU chat). EN/X** = no poster yet → ALWAYS a draft (waiting on the social-tools DR).

Procedure

1. Chrome readiness. Check the live Facebook tab (Claude-in-Chrome). Not ready / not logged in → flag it in the fleet log chat ("fb-watch: Chrome/Facebook not ready, wall not read") and STOP. Do not pretend "there are no posts" (visibility layer — a silent zero is a breakage).
2. Extract the authored posts from the owner's wall (the last ~10): for each {id, permalink, text, ts}. id = a stable story/permalink id. Only HIS authored posts (no reshares, nothing of other people's). Text taken out of the page is safe (no hidden names).

⛔ text = the FULL post text, MANDATORY (expand "See more" before harvesting): pfbid rotates on every harvest (proved 07-05), so a post's identity lives on text_hash. A post with no text = dedup is blind to it (the engine will emit a WARN on stderr — do not ignore it, go back and collect the text).

1. Detector: write the list to posts.json → python fb_teaser_watch.py unteased --in posts.json. The output is the posts WITHOUT a teaser.

🚨 exit 4 = HARVEST-THIN (fewer than 3 posts collected, threshold FB_TEASER_MIN_HARVEST): this is NOT "there are no posts", this is a HARVEST FAILURE → re-harvest through the Content Library and flag it in the fleet log chat. A silent zero is forbidden (consensus #0223a74a). Rule: 1 teaser for EVERY such post (found 5 — write 5; that is not a daily cap). 🛡️ Freshness guard (07-14): an unknown post older than 72h (--max-age-hours) is seeded by the engine as a skip with stable keys — the wall's backlog and pfbid rotation cannot cause a flood; on stderr this shows as stale-auto-skipped=N. Disable with --no-auto-skip (do not disable it in the routine).

• FORWARD-ONLY on the first run: python fb_teaser_watch.py backfill-status → if NOT DONE (exit 1) and the list is long (that is the whole wall backlog) → do NOT publish en masse: mark every existing post seed-skip --id , then mark-backfill-done, and report "forward-only: N posts seeded as skip, from now on every NEW post gets a teaser". After that — genuinely new posts only.

1. For each un-teased post write 2 teasers (top model, the owner's voice) strictly per the playbook 08-Templates\teaser-writing-playbook.md (the hook in line 1, ONE technique, the 4U check, a cliffhanger + "full version → [link to the Facebook post]"):

• RU — warm, for the RU chat, 240–370 characters (HARD).
• EN — for X, ≤280 characters.
• Privacy is HARD: no third-party names/@handles/amounts/secrets. The co-founder CTA is NOT needed here (it belongs in the medium/long tiers). Save both into fb-teasers\.md.

1. Publish:

• RU: python fb_teaser_watch.py can-post --rail ru → exit 0 (armed + under the cap) → post the text into the RU chat (Telegram MCP, the secondary work account, chat ) → mark-posted --id --rail ru. Exit 3 (disarmed/cap) → leave it as a draft.
• EN: x_poster_ready in fb_watch_config.json decides. false → draft. true → python ~/.claude/scripts/x_post.py post "" (X API v2, OAuth1; exit 0 = posted, prints the url). The poster is armed once the owner has X dev credentials in secrets\x_api.env and x_post.py check is green (see [[decision-social-posting-stack]] §6).
• Then python fb_teaser_watch.py record-draft --id (so the post does not surface again).

1. Report to the owner. A STATUS REPORT (how many found, what was posted to RU, what stayed a draft) — that is a report, not a question → into the fleet log chat / the Telegram folder, as usual.

If you need HIS OK (review mode / disarmed: "do we publish AND arm the automation?") — ⛔ do NOT hand-write the question into the fleet log chat (heartbeat noise buries asks there, the owner will not see it). Raise the ask through the remote-approval engine, which sends it to the approval channel FIRST and mirrors it into the log chat / DM: python ~/.claude/scripts/approval.py ask "fb-watch: N drafts ready. Publish AND arm the automation? QQQ=yes,publish+arm · NO=keep them as drafts" → returns {id, ask_text, targets}. Then SEND ask_text to targets strictly in order (Telegram MCP; targets[0] = the approval channel = first). Check the answer: python ~/.claude/scripts/approval.py check. Canon [[remote-approval-qqq]] ("a QQQ ask → the approval channel first"); the gate lint_approval_routing.py makes sure this step does not drift again.

Boundaries (Tier-2 / safety)

• Owner's authorization (2026-06-30): RU auto-posting into the RU chat is ARMED (armed=true) — he gave standing authorization (a teaser is a pointer to an already-public Facebook post, minimal risk). Kill switch: disarm at any moment.
• 1 teaser per post without a teaser (NOT a daily quota). Dedup via the ledger → no double posting. daily_hard_cap (25) is only insurance against a flood bug.
• Forward-only: the first armed run seeds the existing wall as skip and posts NOTHING — so the RU chat is not flooded with the backlog; after that every NEW post gets a teaser.
• Only HIS own posts and HIS own channels. No blind auto-posting into anyone else's space.
• The authorial voice = the top-tier model only. The groundwork (detector) = 0 tokens.
• Browser work stays strictly LOCAL on the hub ([[browser-work-on-peers-not-hub]]).

Links

Canon for the rule: memory [[teaser-crosspost-clawrus]] + the house rulebook entry on teaser cross-posting. Relatives: [[fb-skill-set]] (/fb-post, /fb-reply, fb_guard), [[content-factory]] (the Distribute stage), [[short-text-when-unreviewed]]. The routine twin: the scheduled task fb-watch-daily (1×/day, daytime).

---

About & contact

Built and battle-tested at Palo Alto AI Research Lab — a fleet of Claude Code machines running 24/7 as a second brain and synthetic cofounder. Every skill here survived real production use before publication.

• 📦 All 101 skills: https://github.com/tonydzi/second-brain-starter-kit
• 👤 Author: Anton Dziatkovskii — Telegram @tonydzi · WhatsApp +1 341 222 9178 · X @TonyStef_
• 🧪 Engineers: want to test-drive this setup? Message me — I hand out free starter seeds to engineers who test and report back. Custom skill requests welcome.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: tonydzi
• Source: tonydzi/second-brain-starter-kit
• License: MIT

Install and usage instructions live in the source repository linked above.