Last night for dinner, I made my family this skillet chicken recipe that I found on the New York Times cooking app. But when I first read through the recipe, a few things popped out at me:
It called for a 12-inch ovenproof skillet and indicated that the pan size was very important here. I do not own one.
It called for 2 spicy hot peppers in cooking, but my kids never eat spicy food.
It called for sour cream on the side, but that’s not the healthiest option.
While I was preparing, I decided to make a few minor adjustments to make the recipe work best for me. Here's what I did:
Real-Time Recipe Adjustments
In cooking, like in life, you rarely follow the recipe exactly. You tweak it based on your tools, your family’s preferences, and your own ideas of what will taste best.
This process of adjusting and iterating and taste testing is quite literally the same approach I use when working with AI, specifically when creating custom GPTs. Just like with recipes, building and using custom GPTs isn’t about following one rigid path to get a perfect answer—it's about starting with a base and fine-tuning it to work for you.
Remixing a Recipe: How Do You Iterate?
There are a few “greatest hits” recipes in our family household, and spinach lasagna is one of them. It’s a recipe I co-opted from my parents, one that I had growing up. My mom originally gave me the recipe on a hand-written card. I took a photo of that recipe and saved it to my camera roll.
I’ve probably made a version of that lasagna about 6 times since I’ve had kids myself. Each time I make it, I make 1-2 small tweaks or iterations. First, I subbed out frozen spinach for fresh spinach. Another time, I added in cut-up bits of fresh basil to the fresh spinach mixture to brighten up the flavor. A third time, I cut up an onion and mixed it into the cheesy sauce mixture. Once, I cut up an eggplant and mixed it in with the red sauce. I also really like layering in red pepper from a jar as a middle layer of the lasagna mix.
Each of these tweaks is like refining a GPT prompt—swapping out inputs to see how it changes the output. Just like how I first replaced frozen spinach with fresh spinach to improve texture, in AI, I might replace a generic phrase with more specific language to get closer to the desired result.
Today, I no longer call this dish “spinach lasagna.” It’s straight up “vegetable lasagna.” Spinach is still the base, but through all the iterations and improvements, it’s slowly morphed into something new. Something that suits my cooking style and my family’s eating preferences.
This slow, iterative process of running an experiment, tasting the result, and making improvements is exactly what using LLMs and AI tools invite us to do. With every tweak, you get closer to something that’s custom-made for your needs—whether it’s lasagna or a perfectly tailored GPT.
Riffing on a Recipe vs. Designing a Cookbook
If you use LLMs in any part of your regular workflow, you probably already know that the magic often comes from the follow-up prompts and subsequent iterations, as opposed to the initial output. You can practice this iterative process in refining prompts from a general GPT (like ChatGPT). (ie: "Now give me another recipe, but take out the hot peppers.")
But the major unlock for me in my own usage of AI tools has been in the hyper-personalization allowed from custom GPTs. That’s to say, rather than simply use the same general instance of ChatGPT each time I open the browser or app, I frequently toggle between several micro-apps that I built specific to my own workflow. General GPTs are great for content on a one-to-one basis (like a single recipe or document), but custom GPTs help me generate multiple outputs simultaneously, tailored to different needs or with very specific personalization details. (ie: "Help me plan meals for the entire week for my family, based on our preferences.")
I shared some of this in my post last week, The Three-Body Problem (and Opportunity), but I have built an array custom GPTs suited to my own workflow as a fractional worker. These include things like a proposal builder or a strategy document creator, each trained with basic documentation about my active jobs. This lets me quickly generate first drafts of relevant content using the most up-to-date information.
In other words, a general GPT is really helpful for one-time, in-the-moment triage. A custom GPT is for times when you want to repeat a process (say: plan meals for your family for a week) with the same base criteria. It kind of depends on whether you want to iterate on a single recipe, or help design the cookbook.
Thinking Like a Chef
Creativity through iteration lies at the heart of both cooking and using AI. Just like a chef tweaks a dish to suit their family’s tastes, working with GPTs is about experimenting and making small adjustments until the output fits your needs more precisely. It’s not about seeking the right (or only) answer—it’s about shaping the process to get something that works best for you.
Using a general GPT (like ChatGPT) is great for getting a quick recipe. By contrast, building a custom GPT is like writing the outline for a cookbook, tailored specifically to your unique needs and preferences. Both processes require tweaks and iterations (in how you shape the initial request and in how you adjust the output). But just like standing in your kitchen, refrigerator door open, wondering what to whip up for dinner, the most creative combinations are often right in front of you. You just need to pick a few ingredients to combine together.
So the next time you step into your real or metaphorical kitchen, ask yourself: How can I think more like a chef today?