One Idea. A Blog Post and a YouTube Short. Zero Manual Publishing.
We run a full content operation — blog post plus YouTube Short — from a single idea, with one human approval step and zero manual publishing. Here's the exact system.
We publish twice a week. Blog post plus YouTube Short every time. One person (me) touches the content once — a quick review in Slack. That's it.
Here's exactly how it runs.
The Problem With Content at a Small Agency
You have ideas constantly. You never have time to execute.
You could hire a content writer. You'd spend 30 minutes briefing them, 30 minutes reviewing, 30 minutes editing, then figure out publishing, then figure out the video version. For two posts a week, that's hours gone every single week.
We didn't want a content team. We wanted a content pipeline.
So we built one.
The Full Flow
The Idea Engine
Three times a week, a process called Cassie runs automatically. She's our content AI — built on Claude, connected to a memory system (our "second brain") that knows everything about AzLabs.
Cassie starts by looking for what's already resonating. She pulls data via the YouTube Data API, runs it through Gemini to find outliers — videos in our niche that are punching above their weight. Not the biggest channels. The ones with disproportionate views relative to subscriber count.
That's the signal. If a 300k-subscriber channel is getting 2M views on a specific angle, we want to know why.
From there, Gemini extracts the underlying format — is it a "mistake I made" story? a "vs." comparison? a listicle? — and generates idea variants tailored to AzLabs. We get back a ranked list of topics with the angle, the hook, and the reasoning behind it.
The Draft
Cassie picks the top idea and writes the article.
She reads RAZ-VOICE.md first — a profile of my actual writing style extracted from 165+ emails — before she writes a single word. Short paragraphs. No jargon. Real examples. End with a takeaway. Talk like a smart friend, not a press release.
The draft lands in our vault: an Obsidian vault running on the mini, saved as a proper markdown file with frontmatter — title, slug, date, tags. The whole thing.
Takes about 90 seconds.
The Slack Notification
The moment the draft is saved, a message drops in our #content Slack channel.
It includes:
- The article title
- A two-sentence summary
- A deep link straight into Obsidian (
obsidian://open?vault=...) - Two buttons: Approve and Skip
I open it, read it, hit Approve. If I want changes, I edit the file directly in Obsidian and then approve. The whole review takes maybe 3–5 minutes.
That's my one touch point.
Auto-Publish to azlabs.io
When I hit Approve, a webhook fires.
The publish job runs on our Mac Mini — a $600 machine sitting in my office that handles everything. It copies the markdown into the azlabs.io Next.js codebase, runs a git pull --rebase to stay in sync with origin (so concurrent publishes don't step on each other), builds, deploys, then posts a confirmation back to Slack with the live URL.
About 2 minutes total. I don't open a terminal. I don't touch a deployment dashboard. I just see "Published: azlabs.io/blog/[slug]" in Slack and move on.
The YouTube Short
After the blog post is approved, Cassie generates a 60-second YouTube Short on the same idea — completely different format.
We use Hormozi-style scripts: one insight, structured as hook → the belief the viewer has wrong → the reframe → the receipt. The kind of short that gets rewatched, not skipped.
Cassie writes the script, formats the scene breakdown, and hands it to our video renderer. We use HyperFrames — it takes HTML/CSS/JavaScript compositions and renders them to MP4. The short has animated text, brand chrome, properly-sized typography (48px minimum on a 1080×1920 canvas, so it's actually readable on a phone), and a thumbnail-grade final frame that holds static for a full 1.5 seconds before the video ends.
The whole render — TTS narration via ElevenLabs, animated scenes, final MP4 — runs automatically on the mini.
When it's done, the video gets posted to a private review URL in Slack. Same approval flow: I watch 90 seconds of video, hit Approve or flag a change, and it uploads to YouTube with the title, description, and thumbnail already filled in.
What Actually Makes This Work
The memory system is the thing people underestimate most. Cassie isn't writing generic AI content about "AI trends" or "automation tips." She's writing about our work — real client projects, real outcomes, specific numbers. She can do that because she has access to our second brain: a local Postgres database with hybrid vector + keyword search over our entire vault. Every innovation we've shipped, every client story, every session log is searchable and gets pulled into her context before she writes a word.
The voice profile is the other thing. RAZ-VOICE.md is what most people skip when they try to do this. We extracted it from 165 emails and it's the reason the drafts read like me instead of a marketing intern who's read too many Neil Patel posts.
Then there's the Mac Mini. No cloud. No Vercel. No AWS. Everything runs locally on a machine I already owned. Postgres, the dashboard, the crons, the publish job — all on the same box. We reach it over Tailscale when we're away. It sounds janky but it's the most reliable thing in our stack, and it costs nothing to run.
And the single approval. The reason most content pipelines collapse is checkpoints. Too many people reviewing, too many revision rounds. Ours has one: me, 5 minutes, Slack. If I want to improve the output, I improve the prompts and the voice profile — not the process.
How to Build Your Own Version
You don't need our exact stack. You need four things:
A writing AI that knows your voice. This means actually building a voice profile — pull your last 100 emails or your last 20 posts, extract the patterns, save them somewhere your AI can read before it writes. Without this, everything sounds the same.
A place drafts can land and you can review from. Could be Notion, Obsidian, Google Docs — doesn't matter. What matters is that your AI writes to it and you can approve from wherever you already are (Slack, Telegram, email).
A one-click approval that triggers publishing. Not a five-step process. One button that fires a webhook.
For the Short, you need a script format that works for your niche and a renderer. Hormozi-style works for business content. HyperFrames and Remotion are both solid renderers. ElevenLabs for narration. The script format matters more than the tool.
The part that takes the most time to build is the memory layer — connecting your AI to your own history of work. That's what separates content that sounds like you from content that sounds like everyone else using the same AI. It took us a few months. But once it's there, everything it produces is grounded in actual things we've done and built.
What We're Shipping
Right now: two pieces of content a week, every week. A blog post on azlabs.io, and a YouTube Short on the same topic. Both automated. Both approved by me in under 10 minutes total.
I spend maybe 10 minutes a week on content. That includes reviewing both assets and occasionally fixing a line.
The rest — idea generation, writing, formatting, rendering, deploying — runs without me.
That's the actual goal here. Not "AI helps with content." AI runs content. I'm just the quality gate.
If you want to build this for your own agency or solo operation, start with the voice profile and the single-approval Slack flow. Get those two working first. The rest can follow. The technology isn't the hard part — it's getting the AI to sound like you, not like an AI.
That's what makes people actually read it.
—Raz