The upside-down approach to prompt engineering


Years ago, I invested a significant amount of money and time in creative writing classes.

One month, our instructor gave us a perplexing assignment: write a short story in the style of Ernest Hemingway.

Embarrassingly, I knew who Hemingway was, but I hadn’t read his stuff. It’s all but impossible to write like Hemingway… if you haven’t actually read any Hemingway!

I wasn’t the only one. The instructor, sniffing out our literary ignorance, gave us printouts of one of Hemingway’s short stories and a reading list.

I spent a few weeks devouring The Old Man and The Sea and Hemingway’s short fiction. I even printed a few and wrote them out by hand. After doing this a few times, I began to understand why Hemingway was so good at showing rather than telling readers how a story unfolds.

Then, I tried my hand at the assignment. I can’t say I wrote anything on par with Hemingway, but the results were far better than when I tried the assignment cold.

A few years later, when teaching myself the art of copywriting, I followed the same approach. I found a reading list of the great sales pages and letters of all time. I printed out a dozen by top-grossing copywriters, such as Eugene Schwartz and John Caples, and wrote them out by hand.

The simple act of gathering samples from these writers and writing out their work helped me understand how Hemingway and Eugene Schwartz (creators on different ends of the spectrum) played around with words. I’ve used this approach to learn the story of dozens of writers and creators.

AI critics often point out that the content is bland and generic. They have a point, but if you want to get more value from ChatGPT or Claude AI, flip this approach on its head. Think of yourself as the cranky writing instructor and AI as a fresh-faced student that you’re mentoring.

Most people, when using AI, only pull one lever: the prompt. You’ll get far better outputs if you give ChatGPT or Claude a straightforward assignment (the prompt)… and examples of what you want (personalized training data).

I spend an hour or two each week writing and collecting prompts so I can tell AI exactly what I want. I spend the same amount of time gathering examples of ideal outputs that I can train AI on.

I’m on a Mac, and I write emails like this one in plain text files. I tag these files and upload them to my custom Claude and ChatGPT projects.

Then, when I fire up ChatGPT or Claude, I pull these two levers to get what I need: the prompt and my training data. I can iterate the prompt to provide AI constraints like a writing instructor. And I can give AI clear guidance by showing it, rather than telling it, what I need.

Once you build your training data library, every future chat with AI becomes exponentially easier and faster.

If you need help mastering prompt engineering, check out my newsletter course PromptWritingStudio. It currently costs $25 per month, as it’s in beta. However, it won’t stay in beta for much longer, as I recently completed the final lesson.

In a few days, I’m taking down the "Buy Now" buttons. I’ll relaunch it later this year, but at a much higher price point.

Subscribe to Creator Leverage: Master AI. Build Systems. Grow Your Business