Project · Self-directed · Automation · AI · Media

AI content pipelines: The Ninth Volume, The Ninth Archive, TFCD

Idea → script → human approval → images → narration → render → publish → anthology, as n8n pipelines running on the Pi.

content brands
3
workflows per brand
7
human step: approve
1

Summary

Three content operations run on the same architecture: a weird-fiction story channel, a documentary-style entity archive, and The Food Compliance Desk, a channel showing BRCGS Food Safety Issue 9 workflows with Microsoft Copilot. Seed banks feed idea generators, Gemini drafts scripts, and render services produce long-form video, Shorts and TikTok clips. I approve each Ninth Volume and Ninth Archive script from an email form, and approved stories are compiled into KDP anthologies. TFCD was rebuilt in October 2026 around its own engine, where an automated quality gate takes the place of the email approval.

The problem

Content pipelines break in quiet ways. Approvals time out without anyone noticing, renders fail while the run is marked as a success, and the same story gets regenerated again and again.

Why I built it

To find out how far a one-person content operation can go when the human only does the parts that need taste: picking ideas and approving scripts.

The solution

Every stage writes back to a content plan sheet. Approval is a hard gate. Heavy work goes through the Pi Ops lock, finished voice and render output is reused rather than regenerated, and publishing only marks a row as done after the platform confirms the post.

What it does

  • Seed banks with least-used-motif selection and deterministic tie-breaks, so ideas don't cluster.
  • Gmail send-and-wait approval forms (Approve / Request changes / Decline) with one bounded revision round.
  • Per-scene Gemini images, narration in a cloned voice, and animated renders with camera moves, transitions and captions.
  • Shorts go to YouTube and TikTok through Buffer's GraphQL API, and long-form goes straight to YouTube. A row is only marked posted once a post ID comes back, so a failure retries the next day by itself.
  • Voice and render steps reuse finished output by content key, so a retry carries on from where it stopped instead of regenerating. The audit had found one Short regenerated 17 times.
  • An anthology readiness check queues a KDP print volume once about 200 pages of approved stories exist.
  • The TFCD rebuild: one engine moves each item through ten stages, from research to publish. BRCGS clause numbers and Copilot tier facts come from versioned reference files with sources, not from the model. QC rejects any clause the reference doesn't contain and allows one automatic revision before the item is flagged to me.
  • Every finished TFCD item gets a page on this site with the worked example, prompts and references. The page exists from the end of production rather than from publication, so the link in the video description works on upload day, and a daily n8n check loads every page and emails me if one fails.
  • Found why TFCD had gone silent for weeks: when an approval timed out, the switch node matched nothing, and n8n recorded a success.

Technologies, and what they're for

  • n8nSeven workflows per channel, from ideas to publishing
  • GeminiScripts, metadata, hooks, scene images, thumbnails
  • Chatterbox (voice-render)Self-hosted TTS in my cloned voice
  • Python + ffmpeg render servicesLong-form, Shorts, captions, stingers, end cards
  • Google Sheets + DriveContent plans, script library, asset storage
  • YouTube + BufferLong-form straight to YouTube; Shorts to YouTube and TikTok through Buffer
  • KDP PDF renderingPrint anthologies and handbooks

AI involvement

  • Gemini agents for scripts and metadata
  • Gemini image generation per scene
  • Self-hosted voice cloning (Chatterbox)
  • Microsoft Copilot as the subject of the TFCD tutorials

Automation

  • Nightly idea generation
  • Email approval gates (Ninth) and an automated QC gate (TFCD)
  • Render and polling loops with timeouts and retries
  • Daily Shorts/TikTok publishing
  • Anthology compilation

Infrastructure

  • voice-render, makeit-media, shorts-render and a stitcher, all in Docker on the Pi
  • Moved from n8n Cloud and Google Cloud Run to the Pi

Hard parts

  • CPU starvation: uncapped TTS inference starved n8n's event loop and crashed it. Fixed with container CPU caps and CPU shares, then load-tested under a 40-minute inference job.
  • An 'Expression timed out' error while polling turned out to be oversized items carrying media through the workflow. The rule now is that files go to Drive and only their references travel through n8n.

Outcome

Distribution is live on YouTube and TikTok. TFCD published its first rendered video in July 2026, and since the rebuild its Quick Guides have gone out on YouTube every other day, starting 6 October 2026. Fixes from each pipeline are ported to the others, since all three share one architecture.