My Writing Bot Was Doing an Impression of Me

For the second week in a row, the blog draft was garbage.

I have an agentic workflow that reads my meeting notes, looks for something worth writing about, and sends me a draft. The idea is to give me somewhere to start. I don't want it to write for me and I don't want to just push AI generated content.

I have a lot of examples of my writing as I have been doing this a long time across different industries. I wrote for my theater, I was paid (yes, paid to write!) for satirical work, essays about parenting children with rare-illness, and on and on. So I took the time to build a "lauren-writer" agent. This gives the LLM a way to know how to write enough like me that I'm not staring at a blank page.

If I've been writing for so long, why even bother with this? Great question! A few reasons. One is the exercise of building this agent. Now that I build AI in my day job, it's important that I know what I'm doing. The other is my week is so busy and there are so many small signals to surface on what could be a possible angle or topic and everyone will tell you one of the hardest things to do is continually finding content to write about. So this agentic workflow helps me solve that problem.

Now back to the garbage drafts I was getting. The writing had all these signals that were apparently supposed to make it sound human. “I could be wrong.” “We haven't fixed anything.” A bit of self-doubt, a bigger lesson, then a little disclaimer so the lesson wouldn't seem too confident.

It didn't sound human, it sounded fake. More specifically, it sounded fake in a way that was trying very hard to sound like me. Clearly, something was amiss. Whether it was drift or something else was what I needed to discover.

So I asked an agent to pair with me and take a look at the pipeline. The pipeline isn't just the lauren-writer; there's a helper agent too, one that digs through stored notes when something needs troubleshooting. The voice instructions had taken things from the writing guide and then just AI-ified it. Stupid stuff like “Refuse the tidy win.” This is actually about how I don't write everything is perfect and I always know what I'm doing. So it distilled that and I ended up with fake-me drafts and AI platitudes.

Meanwhile, the writing step wasn't loading the full articles the guide pointed to. It had the rules and some illustrative sentences, but it wasn't reading complete pieces to see how those choices worked together. That seems like a pretty good way to produce an impression of someone. Also, AI getting lazy is a trend but that's a different post.

One might think that just giving the lauren-writer agent more of my work to read was a next step. However, when doing the review and understanding that a sub-agent helper searches stored notes for troubleshooting advice this is signal that more context does not equal better.

The writing bot and the troubleshooting helper had different problems. In a test, the editor caught a made-up timeline and an extra meeting anecdote the writer had added. I mean that's great but I can also do that when I'm editing the draft. This doesn't tell me if the draft itself is usable. In the end, I tweaked and pulled levers. The outcome was a draft that I could work with.

Now we'll see if next week and the week after are also good enough. If not, well, back to tuning. It would be ironic if I end up in a place where I just have an agent giving me headlines and a list of ideas as that is what I used to do on a Sunday morning when looking for content for my next improv post. Sigh... at least it won't be an impression of me.