July 17, 2025 - Flavorful, So Far

I started working on Flavorful for a very pragmatic reason: I wanted to experiment with LLM tools, but I needed to know when they were bullshitting me. Cooking was perfect—I know enough to spot when an AI suggests making Swedish Lemon Angels, but I'm no professional chef with rigid ideas about the "right" way to do things.

I was surprised how good AI is with recipes, but it actually makes a lot of sense. Recipes share a lot of qualities with code—structured but flexible, language-based but grounded in physical reality (either the pudding sets or it doesn't), and extensively documented in LLM training data. But more interesting than AI's competence is the question of what to actually do with that capability.

So far I've built and rebuilt "Flavorful" four times, each iteration teaching me something new:

Version 1: The Recipe Modifier - A Python notebook that transforms recipes based on user input. Want to make that pasta dish vegan? Need to use up some mushrooms? It worked, but felt like a party trick.

Version 2: The Recipe Forker - A full web app where users can save recipes and create "forks" based on modifications. I essentially built yet another recipe database with an AI feature bolted on. The variations were too unpredictable for actual cooking, and I realized: we don't need more recipes. Google already gives us infinite recipes. Creating new ones isn't the problem.

Version 3: The Recipe Annotator - A Chrome extension that analyzes any recipe web page, extracts insights from user comments, explains ingredient roles, suggests substitutions, and recommends pairings. I got great feedback on this one, but the Chrome extension format limited its usefulness. Still, I felt I was onto something—context and understanding matter more than generation.

Version 4: The Shopping List Flipper - My most ambitious version, attempting to reverse the traditional flow. Instead of recipe→shopping list, what if you started with what's in season or what's on sale on your shopping list, and the app suggested recipes using AI-powered "frameworks"? A few trusted testers told me the truth: it was too complex, the workflow felt backwards, and they missed having their own recipes in the system.

Each prototype has revealed deeper questions:

What's the real problem? I keep saying "people get stuck in cooking ruts," but that's vague. Maybe it's more specific: we know 20 dishes we rotate through, we're bored but scared to waste food/time/money on experiments, we see amazing ingredients but don't know what to do with them, we want to eat more plant-based proteins but default to chicken breasts because that's what we're comfortable with.

Where does this tool live? On your phone while shopping? On your laptop while meal planning? In your kitchen while cooking? Is it a traditional app? A voice chat? Something in between? I've been thinking of this as the "surface area" of the app.

How do you build trust? This is easy to ignore but fundamental. If an AI recipe fails, you've wasted food, time, and confidence. One bad recipe could destroy faith in the entire system. Traditional recipes earn trust through reviews and reputation—how does an AI cooking tool earn trust?

I'm exploring three directions right now:

A Recipe MCP Server - Give Claude or ChatGPT access to your actual recipe collection and preferences. These tools are already good with recipes; personal context could make them genuinely great. "What can I make for dinner tonight without turning on the stove?" gets better answers when the AI is aware of your favorite recipes, or what was on your shopping list three days ago.

Recipe as Dialogue - A turn-based conversation for building recipes. You start: "It's hot out and I want salmon." The AI suggests: "How about poke or quick broiled salmon?" You choose poke. It proposes a marinade. You can modify, ask why, learn techniques. Each turn builds understanding alongside the dish. I also think this approach could result in something fun, which is an emotion I'd love to tap into but has been missing from the prototypes so far (think of something riffing on the game Cooking Mama).

Plant-Based Cooking - I'm not a vegan, but I've been trying to eat more plant-based proteins. I've found this transition to be a challenge. I have several plant-based cookbooks, but there's a bit of an "all or nothing" mentality that has made them kind of hard to integrate into my cooking routine. Focusing on this specific problem might answer some of the questions raised by the prototypes so far.

Why do I keep coming back to Flavorful? Cooking is one of the most human things we do—creative, cultural, scientific, necessary. But many of us approach it with anxiety, limitation, or routine. I am not excited about AI cooking for me, or "programming" me by telling me exactly what to cook. But I see a way AI could be a patient teacher, a creative collaborator, and a confidence builder that helps us see our kitchens as workshops instead of factories or ghost towns.

December 9, 2024 - The Grilled Cheese Meditation

I started using the idea of variations on grilled cheese as an example when explaining Flavorful to people, and turned that idea into a short post on my old website:

What is “grilled cheese”? Most of us probably have a specific idea come to mind – for me, it’s Pepperidge Farm white bread and Kraft American cheese, griddled in butter (how my mom made it, of course).

That’s just my own preference – a general definition of grilled cheese might be: cheese between bread, cooked until the bread is toasty and the cheese is melty.

By riffing on this simple formula, the creative possibilities start to unfold. Indeed, the New York Times Cooking site has thirteen recipes for “grilled cheese” that are recognizable twists on our general definition. There are variations on technique (sheet pan, air fryer, grill), type of cheese (cheddar, American, muenster), and extra ingredients (kimchi, sweet potato, onions).

These thirteen recipes are just the tip of the iceberg, each variation sparks others. A kimchi grilled cheese suggests the existence of a sauerkraut grilled cheese. The french onion grilled cheese points to other soup-inspired sandwiches…how about broccoli-cheddar grilled cheese? The toum grilled cheese, incorporating the Lebanese condiment, reminds us there are hundreds of cuisines that might be able to cast their flavors and ingredients onto the blank canvas of grilled cheese. There must be thousands of grilled cheeses, maybe more!

As cooks, how can we see all those possibilities? For most of us, this skill develops slowly over time through study, observation, and trial and error. Even with experience, it's easy to fall into ruts and miss the opportunity for a creative twist or convenient modification. What if we had tools at hand to help spark our creativity? Instead of finding one grilled cheese recipe, or even thirteen, is there a way for us to navigate through the space where thousands of grilled cheese sandwiches exist to find one that's right for this moment?

As an enthusiastic home cook and a curious software developer, I’ve been working on ways to apply generative artificial intelligence to this challenge. In the course of this project, which I've dubbed Flavorful, I've come to see how generative AI can be a lens for exploration. Rather than replacing human creativity, it can be deployed as a collaborative tool that helps us see the familiar in a new way. With this perspective all our recipes, even the simple grilled cheese, become an invitation to play.