The agent wrote the course — Chester Frazier

I gave Cass a job last night: build the course.

I’ve been meaning to write up how I built her for months. The architecture, the hard parts, the failures. Enough people have asked that I want something I can point to. But “write the course” kept slipping below things that felt more urgent.

So I put it on the overnight queue and went to sleep.

This morning I have eight module drafts, seven newsletter issues, a landing page, and social copy for LinkedIn. The repo is at cwfrazier1/build-your-own-cass.

The content is good. Not “decent for something an AI generated” good — just good. Better than I would have written at 2 AM, which is when I would have written it if I were doing it myself.

The obvious irony

An AI wrote a course about how to build an AI that works while you sleep.

I’ve been sitting with this for the last hour and I think the irony is actually instructive rather than just cute.

The course exists because the agent ran overnight. If I had to write it during waking hours, it would have been a blog post series I kept deferring. The overnight slot doesn’t compete with anything — there’s nothing else happening at 2 AM. It’s pure capacity that would otherwise be zero.

But the capacity is only useful if you give it something to do that has a real output. “Research the market for AI tools” produces a document nobody reads. “Write the full course, module by module, and commit each one to GitHub with tests on any code” produces something you can ship.

The output condition is what makes it work.

What I actually reviewed

I woke up at 7 AM and spent 45 minutes reading through what was built.

Things I noticed:

The voice module (issue 7) has a detail I hadn’t written down anywhere — the “Talk soon” exit phrase that makes calls feel less awkward than “Anything else? Anything else?” loops. I said that to a friend once. Apparently the conversation ended up in context somewhere. It’s in the newsletter now.

The capstone (Module 8) covers the job application pipeline architecture — coordinator + worker + SQS — which I built a year ago and never documented. I kept meaning to write it up. Now it’s 700 words of runnable code with explanations.

The landing page is embarrassingly better than anything I would have designed. I would have spent 45 minutes procrastinating on colors.

What required a human

The content itself: no intervention needed. The architecture is correct because the agent has been running on this stack for a year. The code snippets are real code, not pseudocode dressed up as examples.

The things I’ll touch before launch:

Register a domain and set up Substack (can’t do either autonomously)
Post the LinkedIn announcement (could write it, but sending things to the internet while I’m asleep felt like one step past the line)
Review the course pricing (Gumroad listing says $149; I want to think about whether that’s right)

That’s it. Maybe 30 minutes of Chester work to launch something that would have taken me three weekends.

The actual launch

I’m publishing the newsletter at buildyourowncass.substack.com (once I create the Substack account). Issue 1 will go out this week.

If you’re an engineer who’s thought about building a personal AI agent and never gotten around to it, this is for you. The stack is simple, the costs are real ($7–12/month), and by Module 3 you’ll have something running.

Building agents that work while you sleep means sleeping more, not less. I find that satisfying.