← Portfolio  /  ReelRunner
Case Study · Generative Media Pipeline

A video channel that runs itself

One n8n workflow picks the idea, writes the prompt, makes an eight-second video with sound, waits for me to approve it, then posts to YouTube and Facebook. I pointed it at two brand new accounts to see if anyone would watch.

See what it did → Look at the workflow
n8n Google Veo Gemini YouTube API Facebook API
Built, running, and publishing on a schedule.
A glass dragon fruit being sliced A glass plum being sliced A glass strawberry being sliced A glass dragon fruit being sliced A glass plum being sliced A glass strawberry being sliced A glass fig being sliced A molten chocolate shell cracking open A glass dragon fruit being sliced A glass fig being sliced A molten chocolate shell cracking open A glass dragon fruit being sliced A glass strawberry being sliced A glass fig being sliced A molten chocolate shell cracking open A glass strawberry being sliced A glass fig being sliced A molten chocolate shell cracking open Play these below

Real output. Prompt, video, title, description
and upload all written by the pipeline.

The problem

Short-form only works if you never stop

Short video only works if you post every day. Thinking up a fresh idea every day is the part people quit.

Doing it by hand

Every clip is a fresh decision

Pick an idea, write the prompt, generate, watch it, bin most of them, write the title and description, then upload to each platform separately.

  • 20 to 30 minutes per clip
  • Ideas repeat, because nobody remembers what already ran
  • One failure at 11pm and the day gets skipped
What I wanted instead

One decision, not fifty

A pipeline that does all the mechanical work and leaves a person one job: watch the clip, say yes or no.

  • Picks subjects it has not used recently
  • Handles its own failures and retries
  • One approval posts to both platforms
What I built

Six stages, one human checkpoint

The video model is the easy part. Everything around it is what makes it safe to leave running.

Plan Picks a subject it has not used lately
Prompt Gemini writes the prompt and title
Generate Veo makes an eight-second clip with sound
Guard Retries what fails, stops after three
Approve You watch it and say yes or no
Publish One yes posts to both platforms
The workflow

Twenty-one nodes, drawn from the real thing

Names and layout come straight from the real workflow. Green is a decision, amber is where it stops and waits.

Video Creator n8n · 21 nodes
The real n8n canvas for this workflow: the run is planned, Gemini writes the prompt, a JWT is exchanged for a Google access token, Veo generates the video, the workflow waits and re-checks until it is done, then a human approves it and it uploads to Facebook and YouTube.
true false true false true false true Execute workflow Plan Run Generate? Generate prompt Google Gemini Chat Model Structured Output Parser Secret Payload JWT Exchange JWT for access token Generate Video Wait Get Generated Video isDone? isFiltered? isError? Convert to File Review & Approve Approved? Restore Video Upload a video Facebook Graph API

Scroll the canvas sideways to follow the run end to end

The result

Two accounts, started from nothing

New YouTube channel, new Facebook page, zero followers. I let the workflow publish. This is what it made.

35K+ views on YouTube
1.3K views on the top Facebook post
5 days since both accounts went live

As reported by YouTube and Facebook on 28 August 2026, five days after the first upload. Both accounts started from zero and the counts are still climbing.

6 featured clips
Shorts 21.8K views
Shorts 2.4K views
Reels 1.3K views
Shorts 2K views
Shorts 2K views
Shorts 1.5K views

Tap any phone to play it with sound.

Want something like this pointed at your own content?

I build automation that takes repetitive production work off a team without handing away the final say. Open to client work and full-time roles.

Get in touch →

AI transparency

This build is generative by design. A language model writes each prompt and a video model produces every frame, so the clips are synthetic and are published as entertainment, not as footage of real events. No clip reaches either platform without a person watching it first and choosing Publish.

AI tools helped build the workflow itself. The architecture, the guard rails and the testing are mine.