How I Tried (and Failed) to Make $1,000 With ChatGPT

And what it taught me about AI in the real world

Brace yourself: this might be the strangest post I’ve published so far. A few weeks ago, I posed a simple challenge to ChatGPT: “Show me a project you can guide me through that earns $1,000.” Fifty hours, $25 in API fees, a small mountain of half‑working Python and shell scripts, and one extremely dapper steampunk owl later, I can confirm two things:

1. I did not make a thousand dollars.

2. The gap between a cool AI demo and something you can actually ship is wider than it looks—wide enough that investors and builders should make sure they have realistic expectations

Below is a blow‑by‑blow of my adventure, followed by three takeaways for investors and practitioners.

The (Extremely) Short Version

• Goal: Make $1,000 with ChatGPT doing as much of the work as possible.

• ChatGPT’s pick: “Publish an AI‑generated adult coloring book on Amazon KDP (Kindle Direct Publishing).”

• Result: One 40‑page coloring book full of steampunk animals, $0.66 royalty per sale, and a new appreciation for how fragile these tools still are in production

If you want to skip the war story and see the magnificent owl that cost me several weeks of my life, feel free to skip to the last section.

This is where I wanted to include a link to the book on Amazon, but I just found out that Amazon rejected the book because “it contains insufficient bleed.” So, I’m going to have to keep working on this, and I will post the link once I get it live. I’m as disappointed as you are.

Step #1: Generating images

ChatGPT’s idea was quite on trend: cranking out supposedly low‑effort GenAI content that people would pay for.¹ I asked GPT‑3o Pro to outline the steps, and it cheerfully produced a mini‑project plan:

1. Generate 40 black‑and‑white line drawings

2. Lay them out at 8.5″ × 11″, 300 dpi and use Amazon’s KDP tool to generate a book on demand

3. Profit

It seemed so simple. First, I asked it what type of images it thought would sell. The answer was “images of people doing boring things.” Since my goal was to let ChatGPT guide me as much as possible, I went with it. The first image that 4o generated looked okay, but I needed to write a Python script that could generate images automatically. (My idea at this point was that it would be so easy that I would create multiple coloring books with different themes. Ha!)

I ran into many, many issues. I kept getting weird gray bars and artifacts in the images. Predictably, there was some weirdness around hands and feet. It also kept shading in part of the image, which was an issue for a coloring book. Strangest of all, the model kept randomly putting people in Muslim attire. I explained that I didn’t want to be accused of cultural appropriation, but no matter what ChatGPT added to its prompts this never went away. Here’s an example of what I was getting after hours of iteration:

You can see lots of issues here, including table legs in the wrong places and whatever the thing in the lower right is supposed to be.

After a dozen misfires, I pivoted to another suggestion: steampunk animals. Seems like it would be easier than people, and it probably was, but it was still hard.

Some of you may be wondering why the images are so bad, especially since I told you how good the new image model is. It turned out that ChatGPT did not point me to the best image model. It wanted me to use Dall-E 3. Once I used GPT-Image-1,² things improved, but at that point, I was too far down the path on Steampunk Animals to go back to people.

Step #2: Wrestling with the Pipeline

But making the images turned out to be less than half the work. Getting the images into 8.5” x 11” format at the right resolution was very hard. ChatGPT helped me write a shell script to do this. Unfortunately, the script kept breaking and/or failing to produce documents that are the right size. I got tiny animals. I got gigantic animals, but not the right-sized animals.

In the process, I installed Chocolatey (a program just to download other programs), Magick, Potrace, and Inkscape. Getting this script required dozens and dozens of iterations and probably hundreds of back-and-forth prompts. At the end, I couldn’t get the cover image to go from 293 dpi to 300 dpi, which Amazon wanted. After I tried and failed to install another program it wanted called PNG Crush (last updated in 2017). I just decided to submit the 293-dpi version and hope for the best.

I am not proud of this coloring book. This is not what I thought my first work as a published author would be (although it’s not published yet)! Frankly, it stinks, but I just couldn’t take it anymore and hit publish.

Step #3: Profit?

Kindle Direct Publishing (KDP) pays $0.66 per paperback copy at my target price of $6.99. To recoup the $25 I spent on API calls I need to move 38 copies. Hitting $1,000 means selling 1,515 books. If I sell even one, I’ll be very excited. At some point, I will update you on my sales.

Lessons for Investors

What does this have to do with AI and investing? I learned a few lessons that I’ll be taking to my clients:

1. Prototypes are cheap; production is brutal
   A single prompt can generate a demo that looks great. Turning that into a repeatable product is at least two orders of magnitude harder. It’s not impossible, but when I see a company that has launched ANY AI features, I get excited.

2. Image models are still fickle
   GPT‑4o’s image model is miraculous compared with last year’s (as I found out the hard way), but “miraculous” is not the same as deterministic. I needed multiple retries on many of the illustrations. The cover text never got my name and the name of the book correct, no matter how verbose the prompt. If your business model depends on perfect, personalized creative at scale, you have to build in lots of human QA.

3. “AI agents” can’t handle complex tasks
   Everyone is dreaming of autonomous agents that glue services together. My experiments suggest we’re not there. The scripts GPT wrote were reasonable but ran into trouble quickly. If a human following cut‑and‑paste instructions needs 20+ attempts, fully automated end‑to‑end agents are probably not possible with current models

And now the moment you’ve been waiting for. Here is a Steampunk Owl:

Note: The opinions expressed in this article are my own and do not represent the views of Bain & Company.

1. John Oliver’s recent “AI Slop” segment skewers this type of thing. I do draw a distinction between some of the issues he discussed such as people posting fake images on Pinterest and what I’m doing which is selling a product that’s clearly made by AI that people can choose if they want to pay for. And if they choose to buy it, then it means they find value in what the AI created.

2. I’m not sure I need the obligatory footnote about how bad they are at naming things, but here it is.