The reason most AI recipe generators feel useless isn't the AI. It's the prompt. "Make me dinner" is the question; beef wellington with truffle reduction is the answer. You typed the world's vaguest brief into a system that will happily fill in the gaps with whatever sounds most recipe-like, and the gaps include a lot of sherry-glazed shallots.

This article is the fix. A 5-part prompt template, examples, and a method for iterating when the result is still wrong. If you'd rather skip the prompt-engineering and use a system that asks you the right questions automatically, try an AI recipe generator that uses what you have.

Why "make me dinner" gets bad results

LLMs predict the next likely token. When the prompt is sparse, "likely" defaults to "what cookbooks contain a lot of." Cookbooks contain a lot of beef wellington. Cookbooks contain almost no "use these eight wilting things in my fridge before Sunday." So the AI gives you what its training data over-represents, not what you actually need.

The same model, with a precise prompt, will write you the second meal. With a vague prompt, you'll get the first one. The model isn't smarter or dumber; the prompt is.

The 5-part prompt template

Every good recipe prompt has these five parts. Skip any one and the result drifts.

1. Inventory. What you actually have. Not a vibe — the specific stuff. "Half a chicken breast, a wilting head of broccoli, half a yellow onion, eggs, rice, soy sauce, leftover lime."

2. Constraints. Time, equipment, dietary. "I have 25 minutes, a stovetop and a microwave, no oven access tonight, peanut allergy."

3. Preference signal. What you'd ideally eat. "I want something warm and savory, not sweet, not salad-y, with a sauce of some kind."

4. Quantity. How many people. "One adult dinner, one lunch tomorrow."

5. Output format. What you want back. "Give me one main idea, then 2 alternative directions in 2 lines each. No long preamble."

Together, that's the brief. The result will be specific, doable, and won't suggest anything that requires an oven.

Examples (good + bad side by side)

Bad

Make me dinner.

What you get: A formal-sounding recipe for chicken parmesan with a 2,000-word intro, ingredients you don't have, and a cooking time of 90 minutes.

Good

I have: half a roast chicken from yesterday, a head of broccoli, white rice, soy sauce, scallion, sesame oil, eggs. I have 20 minutes, a wok and a microwave. I want something warm, savory, with a sauce. Two adult portions. Output one main idea (3 lines max) and one alternative (2 lines).

What you get: "Chicken fried rice with broccoli — heat oil in wok, scramble eggs, push to side, add diced chicken and chopped broccoli, then cold rice, soy sauce, sesame oil, scallion. Or: same ingredients but stir-fry over rice with a soy-sesame sauce."

The second one is dinner. The first one is content.

Bad

What can I make for breakfast?

Good

I have: 4 eggs, sourdough, butter, sharp cheddar, half an avocado, hot sauce. I have 10 minutes, no time to plate prettily. One person. Suggest two options in 2 lines each.

Bad

Recipe ideas for chicken.

Good

I have 1.2 kg of raw boneless chicken thighs and need 4 dinners across the next 5 days that don't taste similar. I'm okay with some prep time on Sunday. List four meals in one sentence each, with a tag for [batch], [fast], or [showpiece].

You can see the pattern. The information density of the prompt determines the information density of the result.

How to iterate when the result is wrong

The first response is rarely the final one. Treat it like a draft.

| If the response is… | Reply with… | |---|---| | Too fancy | "Make it simpler — pantry staples only, no specialty ingredients." | | Too long | "Cut the intro and the substitutions section. Just the recipe." | | Wrong cuisine | "Make it [Mexican / Thai / Italian / Mediterranean] instead, same ingredients." | | Wrong technique | "I don't have an oven. Stovetop only." | | Missing constraint | "I forgot — I have a peanut allergy. Adjust." | | Suggested an ingredient I don't have | "I don't have [thing]. Substitute or remove." |

Three turns is usually enough. If it's still wrong after three turns, your prompt template is missing one of the five parts above. Go back, add it, restart.

Common mistakes

Mistake 1: Listing ingredients without quantities. "I have chicken, rice, broccoli" reads like an unbounded pantry. Better: "I have 200g cooked chicken, 1 cup of cooked rice, half a head of broccoli." Now the AI can suggest portion sizes that match.

Mistake 2: Asking for "healthy" without defining it. Healthy for who? Lower-calorie, higher-protein, lower-sodium, plant-forward, no added sugar — they're all called healthy and they're all different. Pick one.

Mistake 3: Asking for "easy" without saying easy by what measure. Fast? Few ingredients? Few dishes? One pan? Specify.

Mistake 4: Letting the AI invent. If it says "or you could substitute the kale with collard greens," and you don't have either, it's still inventing. Reply: "remove kale, don't substitute."

Mistake 5: Not telling it what to skip. "No long preamble. No nutritional disclaimer. Just the recipe." This alone saves 200 words of fluff.

FAQ

Can I copy these prompts to ChatGPT or are they only for dedicated recipe apps?

Both work. The 5-part template is model-agnostic. Dedicated apps (like Dietrack's recipe generator) can pull part 1 (inventory) automatically from your kitchen, which saves typing. ChatGPT requires you to type it every time.

How long should a good prompt be?

Long enough to cover the five parts. Usually 50-100 words for one meal. Don't pad — extra adjectives ("delicious," "amazing") waste tokens and don't change the result.

Why does the AI keep suggesting the same dish?

You're probably asking with the same prompt. Add "different from yesterday" or "not chicken-and-rice" to the next prompt. LLMs don't remember between sessions; you have to re-state.

Should I tell the AI my allergies every time?

Yes, in the prompt. Don't trust apps that promise "I'll remember your allergies" until you've verified it across sessions. The cost of forgetting is too high. For ongoing meal planning where the system does track this state, see Dietrack's AI meal planner.