I’ve been using a custom GPT called “Chef Genius” for a while now. It’s a recipe assistant tuned to my kitchen - my equipment, my pantry, my dietary preferences. It generates healthy, vegetable-forward meals using what I actually have on hand.

The value isn’t in the Custom GPT (a MyGPT), but in the prompt that it uses. That prompt works very well in Claude, Gemini, or, even, a local LLM of reasonable capability.

Instead of sharing the prompt and having people edit it by hand, I created an interactive form that generates the prompt automatically.

What It Does

Chef Genius generates recipes that:

  • Use ingredients and equipment you actually have
  • Match your preferred meal style (light, moderate, or hearty)
  • Serve your household size
  • Provide measurements in grams (with US volume for convenience)
  • Include cultural context and prep times

The key philosophy: flavor-first. No sad “healthy” substitutions that ruin the dish. If a recipe needs butter, it uses butter - just in reasonable amounts.

The Template

I’ve created an interactive form that generates a customized prompt for your kitchen. Fill in your equipment, pantry staples, and preferences, then copy the result into any LLM.

Click through for hallucination details and tips.

Hallucinations

Speaking of hallucinations, we’ve all seen the stories of LLMs putting preposterous ingredients in generated recipes or suggesting really dumb ways to improve food. Like glue on pizza or Chocolate Chicken Cake or flat out dangerous suggestions.

Yes, you can make it hallucinate some true bogosity, even dangerous bogosity, if you try. And most of the various horror stories were from people trying. The more subtle risk comes from lack of context. If you fire up the LLM and say “make me a sammich”, there is no context; no clue what parameters, what ingredients, what tools, etc… so, it’ll make stuff up (or, in more recent times, it’ll likely ask clarifying questions).

The key to success with an LLM in any task is to (a) provide clear instructions with lots of context to limit the problem space under consideration and (b) focus on things that the LLM has in its training data.

Given that the models used by the likes of Claude, ChatGPT, Gemini, DeepSeek, etc… were trained on vast troves of internet available data and the internet has about 8 bazillion recipes on it, the recipe construction domain definitely meets (b). By providing a prompt that clearly outlines equipment, ingredients, and goals, that meets (a).

The end result is really quite good at coming up with recipes. Yes, you need to proofread, for sure. But I have yet to see any large LLM (local small models get sketchy fast) suggest anything dangerous, much less even suggesting wildly wrong. The biggest problem I had initially is that the recipes would often be scaled for 1 person or 100, hence the “serves 4 people” guideline.

Tips for Using It

  1. Be specific about equipment - If you have a sous vide or pizza oven, mention it. The LLM will suggest recipes that use them.

  2. List your actual pantry - The more specific, the better. “berbere spice” is more useful than “various spices.”

  3. Update it seasonally - What’s in your fridge changes. Keep a few versions.

  4. Start conversations with constraints - “I have chicken thighs, zucchini, and want something Asian-inspired” works great.

The generator form makes this easier - fill it out once and copy the result.