How We Run the Content Factory Overnight on a Mac

Keep Claude for the two-minute morning check. Put “write article 17 of 45” on a Mac that already sits on your desk. That is the whole idea.

18blog steps you already run

4–7steps that belong on the Mac

1client per laptop per night

You already know the Content Factory. One recording becomes a library. The pain is not the method. The pain is paying a frontier model to babysit the middle of it, then spending the next morning handing the same work to a second chat because the first one burned the budget.

This page is the team SOP for the night shift. Local open-weight models (Qwen and friends) write drafts on the laptop. Scripts fetch transcripts and post WordPress drafts. A named human still owns the client. Nobody needs a new agent operating system to do that.

Keep the expensive model for judgment

The blog posting guidelines are eighteen steps. Claude is currently doing fetch plus write plus publish in one loop because that is convenient. Convenience is also why a 45-episode night costs Opus tokens.

Split the loop.

Step	Who	Internet?	Brain?
List the channel / podcast	Script (`yt-dlp`)	Yes	No
Pull captions	Script	Yes	No
Transcribe if no captions	Whisper on the Mac	After download, no	Speech model
Write the article	Local Qwen	No	Yes
Score the QA checklist	Local Qwen, second pass	No	Yes
Featured image	You	—	Yes
Publish WordPress draft	Script + application password	Yes	No
Flip live, email, Facebook, tag people	You or Claude Chrome	Yes	Judgment

If the local model has to talk to the site, you built it wrong.

Treat the queue like a lunchbox

People worry about agents colliding. You do not need a supervisor swarm. You need a fridge.

The queue is a list of sandwiches. You write your name on one. You make it. Nobody else takes a sandwich with your name on it. If you disappear for four hours, your name peels off.

One client, one laptop, one night. Dennis’s Mac does Sean Kelly. Dylan’s Mac does a different client. They never share a client the same night. That is the whole harness.

The human named Accountable on the Basecamp Project Overview still owns the client. Same parseable labels we already use (Accountable:). The laptop is Responsible. Morning QA is Consulted. Basecamp Updates is Informed. Paying-client truth stays in ClientTracker — do not invent the roster from the public Spotlight directory.

Four boxes. Only the orange one is a language model. The red one stays expensive on purpose.

Run one client folder, not a new product

clients/sean-kelly/
  client.json     channel, site, skip list — no passwords
  voice.md        first or third person, banned phrases
  entity-map.md   people → sites, concepts → our articles
  videos/ID.md    transcript
  drafts/ID.md    article + SEO YAML
  qa/ID.md        pass / fail
  LOCK.json       whose laptop owns tonight

Dinner (15 minutes, no model): inventory, captions, queue rows marked queued.

Night (local Qwen, thinking off): read transcript + guidelines + voice + entity map. Write. Score. Fail once, rewrite once. Fail again, mark needs_human. The agent does not open Chrome.

Morning (you, two minutes per article): voice, broken links, featured image. Then the publisher script POSTs a WordPress draft. Rank Math click-path stays on Claude Chrome when REST cannot set it. Live is a human decision.

Count what a night can actually finish

This is the same pipeline as Trenton, Dylan’s 264, and Ethan’s 45 overnight — with the writer living on disk.

Input	Local time / piece	8-hour night on an M3 Max
8–15 min YouTube	~3–6 min	15–25 articles
45-min podcast	~8–14 min	8–12
2-hour interview	~15–25 min	4–8

Ethan’s 45 short episodes can still be one night. Trenton’s 254 is a week of nights. Claude still wins wall-clock when you are watching. The Mac wins cost, privacy, and “I said this at dinner and it was done at breakfast.”

Sean Kelly’s factory already proved the pattern in waves: 1,680 published guest articles, 790 still in the overnight queue. That leftover is weeks of nights, which is exactly why a queue exists.

Refuse the robot butler

A 27B local model replaces one of your eight to ten agents: the one burning tokens on draft 17 of 45. It does not replace the Chrome floor. It does not replace Dollar-a-Day, the Thank You Machine, Rank Math in the editor, or a whiteboard schematic.

Do not start with Cline. Do not install Hermes to coordinate laptops. Start with one client folder and three commands: fetch at 9pm, write all night, drafts waiting at 7am. That is a VA. Cline is a mechanic.

Thinking stays off. Leave reasoning mode on xhigh and a 45-episode night becomes a 45-episode weekend. Force low/off. One model, one job at a time.

Copy these commands

python3 scripts/status.py
python3 scripts/dinner.py --client sean-kelly --limit 5
python3 scripts/night.py --client sean-kelly --n 5
python3 scripts/morning.py --client sean-kelly --dry-run

The working files live in the Overnight Content Worker folder in this project. README is written for a fifth grader. INSTALL.md is the 20-minute laptop setup (LM Studio or Ollama — pick one). The local model is not installed until a human downloads it. An agent should not silently pull 20 GB onto your disk.

THE DELIVERABLE

Process on the Mac. Promote in the browser. One client per night.

Content Factory Blog posting guidelines

Standing SOP, Aug 17, 2026. Pairs with application passwords, persistent agents, and the receipt for standing this up. Related: the YouTube repurposing pipeline · Trenton’s library · four lunch tickets (which cafeteria pays).

August 18, 2026: every task now has a 1–5 importance bar and sits on the Content Factory line. Agents need a line, not a list.