Learn AI With Me: Writing, Finally

Day 7, in which I finally talk about AI-generated writing

I am the human writer of this post. I’m not an AI, I promise. 100% Human Intelligence. Josh.HI.

For the last few weeks, I’ve been trying to automate what should be a simple task: take a Substack post that I’ve written, use AI to generate a set of Substack Notes, then post them to substack over the course of the week. This turns out to be less-than-simple.

Along the way, I got frustrated trying to create all of the automation. I decided to simply test the core idea—could AI generate a set of useful Substack Notes from one of my posts?

Problem One: AI Doesn’t Sound Like Me

My inspiration for this project was a very good video by Helena Liu, in which she uses Make.com to create a social-media system. At the heart of her system is this prompt:

Act as a social media manager and generate a Facebook post. The post should engage the audience with a compelling introduction, provide essential details, and encourage interaction through likes, comments, and shares. End with a clear call to action.

I decided to put that prompt, slightly modified for Substack, into Perplexity. I gave it the URL to this post, which is about my early thinking about retirement. Here’s what Perplexity returned:

1. On Unexpected Journeys 🚀
Sometimes a simple honeymoon turns into a year abroad! That’s what happened to us—thanks to the generosity of friends and family, our plans changed in the best way. Ever found yourself on an unexpected adventure?
Drop your story below or hit ❤️ if life ever surprised you too!

I suppose that’s credible, but it doesn’t sound even a little bit like me. I don’t use emojis in my posts. I don’t use exclamation marks this way. In short, I don’t speak like a social media manager.

As my friend Greg Battle posted recently, “as with all things, garbage in, garbage out. How will you make sure you’re feeling the AI tools quality input?”

Making AI Sound More Like Me

I decided to improve the quality of my input by asking Perplexity to create a style guide that would make future output sound more like me. Here’s the prompt that I used:

Please look at the three Substack posts, analyze the author's style, then create a style guide that ghostwriters (and AI agents) can use to create future Substack posts in the author's voice.
Here are the posts to use as input….

I have to say that Perplexity did a pretty great job at this. It was kind of uncomfortable to see my writing style—something that I regard as very personal, documented so succinctly.

With the style guide created, I gave Perplexity this prompt:

Please perform this task again, but this time, please generate the Substack Notes using the style described in the attached style guide.

It was a spectacular fail:

3. My “Career Strategy”? Saying Yes, Getting Lucky
Looking back, most of my big moves came from raising my hand for new opportunities. Planning? Sure. Luck? Definitely.
How do you decide what’s next? Smash ❤️ if you’re figuring it out, too!

I have never used the word Smash this way. I never will.

If One AI Fails, Try Another

I’ve come to understand that the various LLMs have different strengths and weaknesss. So I decided to move over to Claude.ai and give that one a try. I’d heard that Claude is better at creating written output. Indeed, I was pleased to see that it did much better, and actually created output that sounds like me.

Claude was not able to read my original post from the URL, so I provided both the post and the style guide as markdown files. Here’s a same of what Claude returned:

Note 1: The Honeymoon Strategy

We blew our wedding money on extending our London honeymoon from 3 weeks to 9 months.

At the time, I wondered if we'd just made a terrible financial mistake. Looking back now? It was an investment in a future I never could have imagined.

Sometimes the best strategy is saying yes to the unexpected opportunity — even when it doesn't make sense on paper.

How have unplanned detours shaped your career?

Full story in my latest post: "What Happens Next?"

Creating Transparency about AI-Generated Content

As I reviewed these posts, I started to get a queasy feeling. Could I really post AI-generated writing and pass it off as my own? I’ve never done that, and I felt like that would be crossing a line that I didn’t want to cross. How much of what we consume is already AI generated? How will we maintain standards if creators are not honest about this?

I decided to augment my style guide with an Ethics section. That section contains the following:

## 8. Ethics

It's important to always distinguish between AI-generated content and human-generated content. For this reason, please follow the following rules:

• If any portion of the generated result has been generated by AI, it must be labeled clearly as AI Generated.

• AI generated content should contain the attribution phrase "This [content/writing/image] was generated by [the name of the tool]."

• in the attribution phrase, substitute the phrases in square brackets for the appropriate value.

It took two attempts for Claude to pay attention correctly to this prompt, but here’s a sample of what Claude returned after it followed my instructions:

Note 4: My Accidental Career Strategy

Turns out I did have a strategy — just couldn't articulate it at the time:

• Say yes to new opportunities

• Follow interesting, avoid boring

• Optimize for learning

• Partner with complementary skills

Not much design to it, but mostly coherent policies that emerged over time.

What's your career strategy been — planned or discovered?

This Note was generated by Claude.

---

Tomorrow, I’ll share another post about writing—this time a comparison of my human-driven writing side-by-side with a piece of AI-created writing on the same topic. The verdict—it’s complicated!

In the mean time, LMK what you think about this piece, and stop by to say hi in the subscriber chat.