Finding recipes in the AI moment we're in

Like many people, I’ve been using ChatGPT to help me come up with recipes to cook for my family. And like most families, we have some specific needs. I am a vegetarian. The rest of my family is not. We have two young boys who have very opinionated about what they eat and get bored very quickly. We worry about whether they are getting enough nutrition, particularly when they hardly touch their dinners.

LLM tools like ChatGPT, Claude and Gemini offer a lot of convenience compared to the more traditional ways we found recipes (websites, blogs and YouTube). We can ask it:

“What should I make for dinner?"
"Can you give me a week of low-sodium meals?"
"What does 100g of salmon give me in omega-3s?”

And they answer instantly.

Actually, what I asked ChatGPT was a bit more elaborate:

“Can you please create recipes for each day of the week for my family with 3 meals a day and 2 snacks. We are a family of 4 with two boys (6 and 2 years old) and the dad is vegetarian. The recipes should meet each family member’s recommended macro and micro nutrient requirements.”

Then I had a back and forth about its recommendations: How do they meet each of the recommended nutrient intakes? What are our recommended intakes? What gaps do we have? What else should we eat to fill those gaps?

What ChatGPT surfaced up sounded great and very healthy. Day 1: Breakfast: Almond Butter Oatmeal with Berries and Chia Seeds… Dinner: “Baked Salmon (or Grilled Tofu) with Steamed Vegetables and Brown Rice”.

I went back and forth with it to understand the nutrient content; it didn’t make sense even in the end. But I moved on and tried their recipes.

I had a couple of winners. ChatGPT suggested a pasta with chickpeas and spinach, which (thanks to San Marzano tomatoes) the kids actually liked. Most of the other recipes, though, the kids did not like. So, I went back to my traditional sources. The recipes my kids did like, I was pretty sure weren’t healthy.

I think it’s fairly clear for most of us by now where LLMs do well and where they fall short.

The ideas at a surface level are great. Sometimes novel and creative. But the recipes aren’t real, they haven’t been made with care, tried on real people and perfected. The nutrition data is guessed at, based on patterns the LLMs have come across before, and obviously not reliable. These are domains where traditional recipes can shine.

But the convenience of LLMs is hard to deny. And that’s where Platebreaker comes in.

The Platebreaker solution

Platebreaker is actually solving a problem I’ve struggled with my entire adult life. Finding recipes that meet my nutritional needs, taste good and are quick and easy to make.

Platebreaker is a search engine that indexes the internet’s recipes and encodes their ingredients with nutrient data (currently solely from USDA FoodData Central). Our search algorithms allow users to find recipes that best close their remaining nutrient gaps, and filter for their dietary preferences and requirements (allergies, ingredients, cuisine types).

When you sign up, we ask a few quick questions (like age and sex) to identify the nutrient targets and upper limits relevant to you. You can also set up profiles for your household so you can find meals that meet the needs of every family member.

Then you can search or browse recipes matched to your preferences and nutrient requirements. The search covers targets and limits for 65 nutrients by default (including essential macro- and micro- nutrients and amino acids) which you can customize and add to anytime for up to 110 nutrients in total. When you add meals to your planner, the search automatically updates for the nutrients you have consumed and stored in your body (See platebreaker.com/docs, if you want to look under the hood for how this works, but the platform is designed so you don’t have to). You can also customize or create recipes, using the same nutrient ranking algorithm used in search to help you find healthy ingredients, and you can use the platform as a tracker to monitor your progress against your targets.

But the real convenience comes when you connect up the platform to an LLM. So instead of searching in Platbreaker you ask Claude “My kids want chicken tonight, what are my options?”. Instead of getting a generic answer, Claude now has access to a search engine of real recipes, wired up with real nutrition data, able to show you recipes best matched to your preferences, nutrient requirements and context for what you’ve already eaten, with links to the real recipes so you can follow their instructions.

When OpenAI rolls out Apps in ChatGPT, you’ll be able to bring up views directly from Platebreaker so you can see and browse the results to your questions from within their platform. Browsing the Platebreaker app when you don’t know what to ask is a great way to get inspiration. Asking an LLM when you have a specific question is fast and convenient. OpenAI’s integrated App experience may offer the best of both worlds and we hope other LLMs follow.

We can think of Platebreaker as the intelligent layer that helps us get more relevant, trusted recipe and nutrient data, harnessing the convenience of LLMs to find and connect us with real content from recipe creators.

Where are we now

We’re in our early phases, building and refining this together with users, creators, and professionals.

• Try Platebreaker: Explore how recipe discovery and nutrition tracking already work today.
• Creators: Visit platebreaker.com/creators and make a free account to our portal control how your recipes are displayed in Platebreaker search results and drive more traffic to your site.
• Professionals: Use the same powerful tools to support your clients. With access sharing, clients can grant you visibility into their meal plans, targets, and tracking. See platebreaker.com/professionals for more.
• Everyone: We’d love you to join and share your feedback with us during our alpha and beta stages.

You can find detailed documentation at platebreaker.com/docs or reach us at contact@platebreaker.com.