Menu Development & Recipe R&D
Menu development is where AI most dramatically changes a chef's workflow. The blank-whiteboard, three-coffees, "what are we doing for spring" session can collapse from a half-day into 45 minutes — without losing any of the creative judgment that makes the menu yours.
The key is to use AI for divergent thinking (generating many options fast) and to use your palate and your team for convergent thinking (deciding which options to test and ship).
What You'll Learn
- A 4-stage AI-powered menu development workflow
- How to generate, narrow, develop, and test dish concepts
- A reusable "recipe R&D" prompt that gives you a draftable starting recipe
- How to use AI to predict failure modes before you waste product
The 4-Stage Menu Development Workflow
Use this workflow whenever you need a new menu, a seasonal refresh, or a single new dish.
Stage 1: Diverge — generate 20 candidate concepts.
Stage 2: Narrow — cut to 6-8 finalists using AI as a critic.
Stage 3: Develop — turn each finalist into a draftable recipe.
Stage 4: Stress-test — ask AI to predict failure modes before you cook.
Each stage uses AI differently. Let's walk through each.
Stage 1: Diverge
The goal here is volume. You want ten times more ideas than you'll keep. Most chefs anchor too early on one concept and never explore the space.
Prompt template:
Act as my menu R&D collaborator.
Context: [restaurant type, average check, audience, location, season]
Constraints: [equipment, food cost target, line capability, allergen mix]
Hero element: [an ingredient, a technique, a cuisine reference, a memory --
whatever sparked this menu]
Give me 20 candidate dish concepts. Each one:
- Dish name (1 line)
- 1-sentence concept
- Hero ingredient
- A "risk note" -- what might go wrong about this dish
Be eclectic. Don't repeat formats. I want some safe and some weird.
Twenty is the right number. Fifteen feels like a brainstorm; thirty feels like a menu archive. Twenty forces variety.
Stage 2: Narrow
Now bring your judgment in. Read the 20. Star the 6-8 you want to take further. Then ask AI to play critic:
Here are the 6 candidates I want to develop further: [paste them]
For each, write:
1. The strongest reason to put this dish on the menu
2. The strongest reason NOT to
3. The most likely guest who orders this dish (one line of profile)
4. Predicted food cost % difficulty (low / medium / high)
Be sharp. I want pushback, not flattery.
This is one of AI's best uses: structured critique. You'll often find that the dish you were most attached to is the one with the weakest commercial case.
Stage 3: Develop the Recipe
For each finalist, generate a draftable starting recipe. This is not the final recipe — it's the recipe you'll cook and adjust in test kitchen.
Act as my recipe development chef.
Dish concept: pan-roasted halibut, smoked corn puree, charred poblano salsa,
crispy hominy, lime-cilantro oil.
Restaurant: 80-seat modern American in Austin, average check $72.
Give me a draft recipe for one portion:
- Ingredient list with weights in grams
- Method as numbered steps
- Mise en place I'll need
- Pickup time (the active minutes during service)
- One "if this fails, the most likely reason is..." note
Use weight, not volume. Be specific about temperatures and times.
Default to techniques a 2-cook line can run.
You will not cook this recipe as written. You will cook a version of it, adjust seasoning, change a quantity, swap a component. That's the point. AI gets you to a 70% draft so your test-kitchen time goes to the 30% that matters: your hands on the food.
Stage 4: Stress-Test Before You Cook
This step is underused and very valuable. Before you fire up the line for R&D, ask AI:
This is the dish I'm about to develop: [paste concept and draft recipe]
Predict the three most likely failure modes:
1. What will go wrong on the line during service
2. What will go wrong with the dish on the guest's table
3. What will go wrong economically (food cost, waste, prep time)
For each, suggest one structural fix.
You'll catch some of these in test kitchen anyway — but catching them before you've sourced 30 lb of halibut and pulled a poblano case from the walk-in is the cheaper way to fail.
A Real Example: A Spring Vegetable Tasting Course
A chef in San Francisco used this workflow for a spring vegetable tasting course around fava beans. Stage 1 gave 20 concepts ranging from a fava bean dashi to a sweet preparation with strawberry. Stage 2 narrowed it to 6, including a fava bean falafel-inspired course and a fava bean ravioli. Stage 3 produced draft recipes for all 6. Stage 4 predicted that the falafel would be too dense without a binder change and the ravioli pasta would be too thick at 1.5mm — both warnings the chef had been about to make. Total elapsed: 90 minutes. Estimated time saved versus the chef's normal R&D cycle: about 4 hours.
A Note on Originality
A reasonable worry: if every chef is using AI for menu R&D, won't every menu start to look the same?
In practice, no — and the reason is exactly what we covered in Lesson 2. The chef who feeds in their specific audience, equipment, location, voice, and constraints gets very different output from the chef who types "spring tasting menu". AI amplifies whatever specificity you bring. Generic input gets generic output.
The chefs whose AI-assisted menus look generic are the ones who weren't being specific to start with. Your specificity is your originality.
Key Takeaways
- Use AI for divergent menu thinking (generate many) and your palate for convergent (decide which)
- The 4-stage workflow: Diverge → Narrow → Develop → Stress-test
- 20 concepts is the right divergence number — variety over volume
- Ask AI for structured critique on your finalists
- Predicted failure modes save you cook-time waste
- Your specificity is your originality — generic prompts give generic menus