Skip to content
bpBrandon Perez
All projects
In DevelopmentAI Agents / Content Automation

YouTube Content Creation Automation

An AI-agent pipeline for short-form video, covering daily trend scouting, AI-assisted production, and approval-gated publishing to YouTube, Instagram, and TikTok through their official APIs.

Last updated Oct 17, 2026

Problem

Problem being solved

Running a short-form video channel is a daily grind: find what's trending, turn it into an idea, produce the video, write captions, post to three platforms, and check what worked. Most of that is repeatable. The goal is to automate everything except the creative call and the final "post it" decision.

Architecture

content-pipeline.arch
  1. 01Trend scout agentdaily research · idea pitches · scheduled job
  2. 02ProductionAI video · voice · captions · ffmpeg assembly
  3. 03Human approvalpreview on phone · nothing posts without it
  4. 04Publishing APIsYouTube Data API · Instagram API · TikTok
  5. 05Analytics loopweekly metrics feed the next ideas

Overview

I'm building this together with an AI agent that runs on my own machine. A team of narrow agents each handles one job: find ideas, produce, publish, measure. The final decision always stays with me.

How it works

  1. Scout: every morning, a scheduled agent researches what is trending and pitches a short list of post ideas with hooks, captions, and a cost estimate.
  2. Produce: I pick an idea. AI tools generate the shots and voice, and a local ffmpeg pipeline assembles the vertical video, captions, and cover frame.
  3. Approve: the video is uploaded to each platform unpublished, and I get a preview on my phone. Nothing goes public without an explicit 👍.
  4. Publish: after approval, it goes live through each platform's official API with platform-specific captions and AI-content labels.
  5. Measure: a weekly report shows what got shares and follows, and that feeds the next round of ideas.

Design principles

  • Official APIs only. No password sharing, scraping, or browser automation on accounts. Each platform is connected through its own developer program and OAuth consent screen.
  • Human in the loop. Upload and publish are separate steps, and publish requires approval.
  • Secrets stay local. Tokens live in a local secrets file and refresh automatically. They are never printed, logged, or sent through chat.
  • Budget guardrails. AI video generation has a monthly cap, and the first shot is always tested before generating a batch.

Lessons learned so far

  • Developer onboarding is most of the work. Business/creator account types, tester roles, consent screens, and privacy-policy requirements take longer than the code. That's one reason this website exists.
  • "Upload by URL" APIs need a reachable file. Free temporary file hosts failed (outages, bot redirects to HTML pages). Serving the file briefly from my own machine through a short-lived tunnel worked reliably.
  • Fresh hostnames and local DNS caching don't mix. Check reachability through public DNS, not the machine's resolver.
  • Center crops ruin vertical video. Reframing shot by shot, and letterboxing group shots over a blurred background, keeps every subject in frame.

Related notes