Three down, one to go. So, workflow 3 is finally where we get to the actual writing. It is fairly straightforward, but not without some questions to be answered. A general run down of workflow 3 is the agent grabs topics that are now in a “verified” state, meaning they were approved as ideas, then again, verified as being factually accurate. So when the status of any line reaches verified, it gathers it up, and runs it through the rest of the workflow.
Again, I have this set to run every 3 hours instead of weekly like the idea generator, so I can approve ideas throughout the week as I have time and focus, and they can move along before the next round of idea emails comes out. Swamping my inbox is just going to make keeping up with it overwhelming, which will lead to inconsistency, which is the opposite of what we want from this agent.
A Note on How AI Agents Actually Work
So, once the workflow picks up the verified topic, it takes it to the AI model. Now, this is where building an agent is genuinely different from the way most of us use AI. When you open a chat interface and type a question, you are using what is called a user message — a specific request for that moment. That is the only message most people ever write. But when you build an agent, you also write a system message: the AI’s standing instructions. Think of it as a job description that never changes between runs. Mine tells the AI it is a staff writer for The Educating Entrepreneur, that the audience is first-time business owners, that the tone should be warm and jargon-free, and — critically — that it must build the post from the verified research Workflow 2 collected, not invent its own statistics or quotes. The user message then supplies the specifics for this particular run: here is the topic, the angle, and the research. Write the post.
The other thing an agent needs is structured output. When you chat with AI, you get a text response back and you read it. But an agent needs to know exactly where each piece goes. So instead of asking for a blog post, I ask for a JSON response — a structured format with labeled fields: title, body, excerpt, tags. The workflow reads each field and automatically drops it into the right column in the Google Sheet, no copy-pasting required.
What the AI Builds — and the Edit Loop
So once through the AI model, the agent places the title, body, excerpt, and tags, all things we need for posting to a blog, into the Google sheet, and uses those fields to help write yet another email for me to review. The email renders the full draft — title, body, excerpt, and tags — so I can read the complete post right in my inbox, even on my phone. Instead of just Approve or Reject, there is a third option: Request Edits. I type my notes right into the form (“the intro is too long,” “add a more concrete example in section two”), hit submit, and the workflow automatically passes those notes back to the AI model along with the previous draft. The AI revises specifically what I flagged — not from scratch. A new email arrives with the updated version, and I can go as many rounds as I need until the draft is right and I hit Approve.
What’s Next: Workflow 4
Once I approve it, the status in the Google Sheet changes to “ready.” That single word is all Workflow 4 is waiting for. Every three hours, Workflow 4 wakes up, checks for anything marked ready, and pushes it straight to WordPress as a draft post — with the title, body, excerpt, and tags already filled in. All that is left is a final visual check and one click to publish. That post is coming next.