COG Content Factory Skill

Purpose

Act as the user's autonomous content creator: scout what just happened in their field, pick what is genuinely trending, and turn it into content in their voice — with zero duplicates, hard volume caps, and verification before anything counts as published. Designed to run unattended on a schedule (nightly) and to be safely re-runnable: an empty run is a valid run; a low-quality post is not.

When to Invoke

• /content-factory / "run the content factory"
• Scheduled runs (e.g. 2-4 times a night via cron/launchd)
• "Turn today's news into content"

Prerequisites

Reads the user's voice and beats from 00-inbox/MY-PROFILE.md and topics from 00-inbox/MY-INTERESTS.md. Publishing targets come from 00-inbox/MY-INTEGRATIONS.md — skip disabled channels silently.

State Files (read FIRST, update LAST — this is the dedup backbone)

• 04-projects/content-factory/ledger.md — everything ever covered/published (topic slug, source URL, formats, links, date). NEVER cover a topic or source URL already in the ledger unless there is genuinely NEW news (a follow-up release counts; rephrasing the same launch does not).
• 04-projects/content-factory/tonight-<YYYY-MM-DD>.md — tonight's run log (create on first run of the night). Records what earlier runs scouted, parked, produced, and published, plus the remaining volume budget.
• 04-projects/content-factory/sources.md — the watchlist. Add sources you find valuable; prune dead ones.

Volume Budget (across ALL runs of one night, hard caps)

Channel	Max/night	Notes
Short social posts/threads	2	the volume channel; a thread counts as 1
Blog post	1	only when substance justifies it
Short video	1	only if the idea is visual/demo-able
Long-form professional post	1	only for a strong professional angle
Long video / deep essay	0	never autonomous; propose in the run log

Track the budget in the tonight file. Later runs inherit what is left. Adjust caps to taste — the invariant is that caps exist and are enforced across runs, not per run.

Phase 1 — SCOUT (every run)

1. Web-search for last-24h announcements across the user's watchlist topics, plus anything already trending.
2. Fetch the watchlist blogs/changelogs in sources.md.
3. If a browser integration is available, skim the user's social feeds for engagement signals — raw engagement numbers are the best trend signal.
4. Timebox scouting to ~20 minutes of work.

Phase 2 — TRIAGE (score, don't vibe)

Score each candidate 1-5 on:

• (a) trend momentum — are major accounts and aggregators talking about it NOW?
• (b) beat fit — does it fit one of the user's beats (from MY-PROFILE.md)?
• (c) unique angle — can we add something ONLY this user can add (an experiment, a practitioner take, a contrarian read) — not a summary anyone could write?

Sum ≥11 → produce. 8-10 → park in the tonight file for a later run to re-score. <8 → ignore.

Dedup check before producing: topic slug + source URL against the ledger AND against the user's recent published posts (last ~20) if reachable.

Phase 3 — PRODUCE (format ladder — decide by substance, not ambition)

• Default: one short post or thread — hook line, media attached, link last, minimal hashtags.
• Blog post only when there are ≥3 original things to say or a real proof-of-concept with artifacts. If the user has a PoC beat: actually build the tiny PoC (timebox ~45 min) and publish the REAL result, including failures. Never fabricate PoC outcomes; "it broke here" is good content.
• Short video when the idea is visual or demo-able in ≤60s.
• Cross-link everything (video ↔ blog ↔ thread reference each other).

Phase 4 — PUBLISH (safety gates)

1. Environment gate: if the publishing surface is unavailable, or the user appears to be actively using the browser/machine — DO NOT publish. Write finished bundles to 04-projects/content-factory/out/<date>/ and let a later run (or the user) publish.
2. Post-condition check (mandatory): after every publish, observe the artifact — screenshot the live post, curl the live URL — before recording it as published. Never report success from the publish call alone.
3. Never engage in arguments, never quote-dunk, never reply to strangers. The factory only creates original posts on the user's own surfaces.
4. Any screenshot must pass a credentials check (no account emails, tokens, or internal URLs visible).
5. Never delete an already-engaged post; edit in place if a fix is needed.

Phase 5 — LEDGER + LOG (always, even for empty runs)

• Append published items to ledger.md: | date | slug | beat | formats | source URL | published links |.
• Update tonight-<date>.md: scouted list, scores, parked items, budget left, what the next run should look at first.
• If a big announcement deserves long-form treatment the user should shape, write it as a proposal in the tonight file with a suggested outline — do not attempt it autonomously.

Voice Checklist (delete the draft if any fail)

• Sounds like a builder sharing, not a news bot ("I tried", "I broke", "here's what surprised me")
• Contains at least one thing only this user could say (their setup, their numbers, their PoC, their lens)
• No AI-slop vocabulary ("game-changer", "revolutionary", "unleash"), no emoji walls, hashtags within caps
• Media attached; link present; nothing confidential
• Not a duplicate (ledger + recent posts checked)

The Loop (see /loop-engineering)

Each night is a scan-until-budget-dry loop across runs: the tonight file is the shared loop state, the ledger is the dedup verifier, the volume budget is the hard cap, and the voice checklist is the quality gate. Termination: budget exhausted, nothing scores ≥11, or the environment gate blocks publishing.