Faceless YouTube automation · operating model · 11 October 2026

How the system runs itself

A pipeline Chief of Staff agent on the VPS conducts every run, reviews each step, reports in its own Slack channel, answers you in plain language, proposes fixes for your approval and keeps a record of what it learns. Three videos a week go public without anyone seeing them first. Click any box to open it.

This pipeline is a standalone system beside the cockpit. It is not routed through the shared coordinator, is not a registered cockpit capability and is not in the brand registry. This page describes it as built on 11 October 2026 and contains no credentials.

Channel
Myth Fine Print @MythFinePrint · mythology · model: Mythological Man
Videos public
3 Slavic 5 Oct · Hindu 9 Oct · Celtic 9 Oct
Rhythm
3 videos a week Mon, Wed, Fri at 09:30 Brussels time
Talk to it
#youtube-pipeline every report, decision and alert; reply in plain language

Who starts what

Read left to right, top to bottom.
Control layerdecides, reviews and reports; does none of the work itself
timerswake the agent
  • Monday 09:30: research, then a video.
  • Wednesday and Friday 09:30: a video.
  • Every day 08:40: health check.
  • Every minute: a root job reads your Slack answers and messages and wakes the agent for them.
  • Each timer tells the agent what kind of run it is, so it never guesses from the date.
Pipeline Chief of Staffconductor, reviewer, voiceproven

Claude Code on the VPS under its own restricted user. It holds no keys. It is not the cockpit's Chief of Staff coordinator, and it is separate from the Cold Take pipeline agent.

  • No script runs on its own clock: the agent asks for each step, and is woken again when the step ends to review it and decide what runs next.
  • One retry per production day, then it alerts and stops.
  • Reads a status snapshot and a read-only copy of the scripts, the latest logs and the last eight productions.
  • Proven: unattended runs on 7 and 9 October, a parked video raised correctly, a failed step diagnosed without a retry.
dispatcher and applierthe agent's hands, run as root
  • Carries out only named requests: the four runs, Slack posts and replies, channel statistics, a status refresh, questions about parked videos, change proposals.
  • One run at a time, daily caps on every action.
  • Applies a change proposal only on your own approval in Slack, with a fingerprint check and a backup.
  • Acts on a video only when your own message asks for it: the agent must point to that message, from the last 6 hours.
#youtube-pipelineone channel for everything
Report
Each published video, the Monday week summary with performance and the playbook review.
Decision needed
Parked videos and change proposals. Answer in the thread.
Alert
A failed step after its retry, low balances, a Claude limit, a refused request.
Chat
Write anything; it answers in the thread.

If the agent itself fails, a fallback posts that its run failed and sends any waiting alerts raw.

Research runMondays · scan the market, keep the model channel current
discoverscan niches and channelsproven

Scans YouTube niches and channels with the YouTube Data API and scores them.

Proof
Run of 5 Oct: 311 channels, 5,860 of 10,000 daily quota units.
notion_syncresults into the Notion hubproven

Writes niches and channels to the Notion hub, so you can see if a stronger niche appears.

Proof
316 channel rows in Notion.
choosekeeps the model channelproven

Picks the strongest niche first, then the best channel in it. Once a channel is chosen, later runs keep it. It only changes if you ask for a re-choice.

Proof
Chose Mythological Man. Rejected a Hindi channel dubbed into English.
analysemodel's uploads rankedproven

Ranks every upload of the model channel against its neighbours and pulls transcripts. Picks up its new uploads each week, so topic picking knows what is already covered.

Proof
Videos database in Notion with outlier tiers.
dnathe format recipeproven

Extracts the title, script and thumbnail recipe of the model channel. Done once per model channel, not refreshed weekly.

Proof
Channel DNA pages in Notion.
↓ feeds every script: the format recipe of the model channel is what the write step follows.
Produce runMonday after research, Wednesday, Friday · one video per run
writetopic, script, fact-checkproven

Picks a topic, researches at least 40 sourced facts, writes and fact-checks the script, then writes title, description, tags and the stamp words for the thumbnail.

Proof
Slavic monsters script: 85 claims checked, 6 fixed, 0 open.
Gate
Unsupported, wrong or legally risky claims park it as Needs review.
rendervoice, images, videoproven

Voice-over, one image per scene, video, captions, thumbnail, then three rounds of visual checks. The thumbnail's red stamp is drawn by code, so the image itself carries no lettering.

Proof
Three videos rendered and published; the stamp fix of 9 October passed on the Hindu redo.
Gate
A thumbnail that fails three times parks the video as Needs review.
publishthrough Blotato, publicproven

Uploads video and thumbnail through Blotato as public, corrects chapter times, updates Notion and Slack.

Proof
Celtic video published fully unattended on 9 October, with thumbnail and tags.
Gate
Refuses anything not marked Ready, and refuses to post the same video twice.
tagsthrough YouTube's own APIproven

Sets the video's tags right after publishing, because Blotato has no field for them.

Proof
26 and 30 tags read back on the Celtic and Hindu videos.
Approval
The Google project is In production since 11 October, so the approval no longer lapses after 7 days.
YouTubeMyth Fine Print3 videos public

Phone-verified, icon and banner on. Videos are marked as containing AI-generated media and not made for kids.

Repair runs: finish (render and publish) and publish (publish only). The agent may use one of them as its single retry.

Talking to it

Everything happens in #youtube-pipeline.
Chatplain language, any time

Write a message or reply in any thread. Within a minute it reacts with eyes and starts; it answers in the thread.

  • Answers from the status, the scripts, the logs, its notes and the playbook.
  • Fetches YouTube numbers for our channel or any other, refreshes balances, searches the public web.
  • Starts research or a video when you ask, also outside the schedule (daily caps still apply).
  • Publishes, redoes the thumbnail of, fixes or drops a parked video when you ask.
Parked videosone-word answers

A video stopped by a check comes to you as a question with the reason.

Rendered
publish, redo thumbnail or drop; react ✅ to publish, ❌ to drop.
Not rendered
fix (the writer fixes the open fact-check points, then render and publish) or drop.
Change proposalsnothing changes without your ✅

When something fails in a way that could happen again, the agent finds the root cause and posts the fix as a diff with a fingerprint.

  • ✅ applies it, with a backup. ❌ rejects it.
  • undo in the thread within 7 days rolls it back.
  • Only the pipeline scripts, the channel identity and the agent's own instructions can be changed this way; the files are locked against the agent itself.
Proof
Test proposal applied on 11 October; an out-of-scope proposal was refused before posting.

How it learns

Every run starts by reading these.
Ledgerone line per run

What woke it, what it did, what failed, what it handled. Never edited afterwards.

Open issuesuntil they clear

Every unresolved problem stays on this list until it clears, however old. An issue already on it is not news: it gets a daily reminder with what changed, not a fresh alarm.

Playbookdated lessons

One dated line per confirmed root cause or confirmed fix, in Operations and Content. Started with five lessons from the first week. Reviewed every Monday: wrong or outdated lines are removed or rewritten, and the summary says so. You can add a lesson in chat.

Performanceour own videoscollecting

Four times a day: views, likes and comments of every published video, with its topic, angle, title, length and thumbnail, at 48 hours, 7 days and 28 days.

Now
Numbers only in the Monday summary. Content lessons start at 12 published videos.
Next
After three Content lessons the agent offers a proposal that feeds them to the writer; it only writes it if you say yes.
Missing
Click-through rate and retention need YouTube Analytics, which is not connected.

The week

Brussels time.
Monday 09:30research, then one video

The agent starts research. When it ends well, it starts one video. If research fails even after its retry, the video is still made with the existing format recipe, and the alert says so.

Wednesday 09:30one video

The agent starts one video. A video parked earlier in the week does not block this one.

Friday 09:30one video

The agent starts one video.

Every day 08:40health check, starts no run

Balances, YouTube quota, Google approval, disk space, open issues and alerts the scripts handed over. Silent unless something needs attention. On Mondays it posts the week summary: videos, performance, balances, open items and the playbook review.

Every 10 minutes · every minute · 4× a dayroot jobs, no Claude usage

Status snapshot and read-only source copy every 10 minutes; Slack answers and messages every minute; performance snapshot at 02:30, 08:30, 14:30 and 20:30.

The agent's limits

May dowithin its caps
  • Conduct the runs. Research, produce and the two repair runs, one retry a day.
  • Act on your word. Extra runs and video actions, only when your own message asks.
  • Investigate and propose. Root cause from the logs and scripts, fix as a proposal.
  • Learn. Ledger, open issues, playbook.
  • Wait out a Claude limit, then retry after the reset.
May not doeven when asked
  • Change a file directly. Only an approved proposal does that.
  • Touch keys, timers, Slack routing, the dispatcher or the applier.
  • Unpublish or delete a video.
  • Send anything out to a website; web access is read-only.
  • Touch the backup or any other brand.

Paid services

Balances on 11 October. Open a box for what it does and when the agent alerts.
OpenRouter$28.51 left
Used for
Scoring topic candidates and writing scene image prompts with DeepSeek V4 Flash. Falls back to Claude Haiku.
Alert
Below $5.
Note
Other systems of yours use the same key.
Claude subscriptionlimit hit once5 Oct, 03:35
Used for
The agent itself (scheduled runs and chat), research, script, fact-check, packaging, visual check.
Alert
When a limit is hit. A percentage cannot be measured on the server, so there is no 90% warning.
Blotato5,773 credits
Used for
Uploading to YouTube.
Alert
Below 1,000 credits.
Apify$7.03 of $19this month
Used for
Transcripts and the spoken-language check.
Alert
At 90% of the budget.
OpenAIabout $2.35 a video
Used for
Voice-over and images. Roughly $7 a week at three videos.
Alert
When it runs out.
YouTube Data API10,000 units a day
Used for
Research scans, tags, performance snapshots and channel statistics on request.
Alert
At 90% used in a day.

Open items

Chat and proposals are newfirst real use this week

Built and tested on 11 October; the rollback (undo) test and the first real chat are still to come.

Too few videos to learn from12 needed

Content lessons start at 12 published videos, about four weeks at three a week.

System disk nearly full4.1 GB free

The agent alerts below 3 GB. The pipeline and the agent live on the data disk and the log disk; only the small timer definitions sit on the system disk.

No human look before publicby design

The fact-check, visual check, thumbnail check and Ready rule are all automatic. A video that passes them goes public unseen.

Video quality gapsmusic, in-video style

Videos are voice and images only, with no background music. The model channel's in-video style was never analysed, because video downloads are blocked from the VPS.

Backup keptnot part of any run

Bagholder Association as a second model channel, with its own identity (The Last Receipt), one rendered Starbucks video and no YouTube channel.