Part 8 of the Heist Build Log — an honest, running account of building a content OS from scratch. New entries drop on the 1st and 15th.
Seven entries in, and I finally did it
After seven posts about why I built Heist, what I broke, what I rebuilt, and how I wired the whole thing into ChatGPT's MCP layer — I handed the keys over.
This entry was planned, written, and scheduled by the system I built.
I'm not sure how to feel about that. So let me tell you exactly what happened.
What I had to get right before I could let go
I'm not the kind of person who automates something and just trusts it. That's probably why building Heist took as long as it did.
Before I could hand the build log to an AI agent, I needed a few things to be true.
The Brain had to actually know my voice. Not just a style guide — it had to have absorbed enough real writing that the output didn't come out sounding like a corporate press release with my name bolted on.
The topics had to be structured. I couldn't automate the judgment of "what's worth writing about." So I wrote a brief: the existing entries, the themes already covered, the rule that each post had to carry the story forward and not just repeat what I'd already said.
And the system had to be honest about the gaps. This was the hard part. A lot of AI-written content sounds confident because it was trained on confident text. I didn't want that. I wanted something that sounded like me when I'm uncertain — which is often.
The uncomfortable thing it revealed
Here's what I didn't expect: the hardest part wasn't the writing. It was defining what "the story so far" actually was.
When I sat down to brief the system, I had to articulate the arc of seven posts in a way a non-human agent could follow. Doing that made me realize I'd been a little all over the place. Entry 3 and entry 5 both basically said "here's how the Brain works." Entry 6 was about pricing but also kind of about the Brain again. The publishing engine entry blurred into the Autopilot entry.
The machine didn't have an opinion about any of that. It just needed a brief. And writing the brief forced me to see what I'd actually written — not what I'd meant to write.
That's what I mean when I say that running your own content through a system teaches you things. It's not that the AI is smarter than you. It's that putting your thinking into a format a machine can follow requires a clarity that your own head never demands.
What the last two weeks actually changed
The MCP connector — covered in the last entry — has been live and getting real use. A few things I've noticed since:
- Tool count matters more than I thought. When you're calling Heist from inside ChatGPT, the cognitive load isn't about how many tools exist. It's about how many you have to think about to accomplish any given thing. We have flows that require four or five separate lookups to do what should be a single step.
- Date math is a quiet footgun. Every agent session I've run, computing "next Tuesday" goes sideways if the current date is Tuesday or close to the boundary. Known issue. Not fixed yet.
- Error messages are too polite. When something fails — a missing field, a scheduling conflict, a permission problem — the tool returns a helpful-sounding response that doesn't say what specifically went wrong. When a human reads it, it's frustrating. When an AI agent reads it, it's useless. Vague errors produce confident wrong guesses.
All of this is fixable. All of it is on the board. The point is that you don't find these things until the system is running in the wild, under real conditions, by callers who aren't you.
What it actually feels like when the system runs itself
Honestly? It mostly feels fine. Which is either a good sign or a bad sign depending on your disposition toward these things.
The moment I was most anxious about was the first time Autopilot sent social posts without me reviewing them. Not the build log — the regular content. Six platforms, scheduled six hours out, no human eyes. The stuff went out. It sounded like me. Nobody complained.
A few posts did better than usual. One person replied that my content had felt "more direct" lately. That was the Brain doing its job — it had picked up that my recent posts were running longer and softening their edges, and it quietly corrected course without being asked.
I noticed. Nobody else needed to.
Why I'm still writing this
This series exists because I believe in building in public. Not the performative version — not the "here's my MRR dashboard" version — but the actual one, where you say what you're wrong about and what you're still figuring out.
Heist is a system that helps people get content out of their head and into the world, consistently. That's it. Building it has been a two-year exercise in being honest about how hard "consistently" actually is — for me, for the product, for the people using it.
The build log will keep running. Automated when it can be. By hand when it should be. I'll tell you which is which.
Try it yourself
If you're spending 90 minutes a day on content and getting inconsistent results, Heist is built for exactly that problem. Seven-day free trial, no credit card required. The Brain learns your voice in the first session.