TL;DR
- Paste your real fridge contents into ChatGPT (GPT-5.5), Claude (Sonnet 4.6), or Gemini 3.1 Pro and ask for one 25-minute dinner using only what you listed. Use the copy-ready prompt below.
- Faster than typing: snap a photo of the open fridge and upload it. All three default models read images and will inventory what they see before suggesting a dish.
- The two things AI gets wrong every time: it overestimates how fast you chop, and it quietly adds an ingredient you do not own. Re-read the ingredient list before you start cooking.
- Taste at the end, salt by the teaspoon, and never trust AI for medical allergens (celiac, nut, shellfish) — verify those ingredient by ingredient yourself.
- For a no-prompt option, free dedicated tools like SuperCook and DishGen match recipes to a pantry list with no signup.
The task
It is 7 PM on a weekday. The takeout app is open in your other tab and you are 90 seconds from giving up. The fridge has chicken, half an onion, an aging bell pepper, a small carton of yogurt, rice in the cupboard, and the usual pantry oils and spices. You do not want to go shopping, you do not want to read a 2000-word recipe blog with eight ads, and you definitely do not want a recipe whose third bullet starts with “buy 3 more things.” You want a 25-minute dinner that uses what you have, tastes like food, and does not require a single judgment call you do not already know how to make.
Where AI helps — and where it does not
AI is genuinely good at combining a random ingredient list into a plausible recipe: mapping what you typed onto a familiar dish skeleton (one-pan stir-fry, sheet-pan bake, skillet-and-rice), then sequencing the steps so things finish at the same time. It is also useful for seasoning logic — when to add acid, when to taste, where salt goes wrong.
What AI cannot do reliably:
- Cook times for your specific stove. Stove power, pan material, altitude, and how cold your protein started all shift cook time by 20-40%. Treat any number as a starting point and check doneness, not the clock.
- Salt amounts that match your salt. This is real, not pedantry. Morton kosher salt is about twice as dense as Diamond Crystal — roughly 1.2 g vs 0.7 g per 1/4 teaspoon, a 2.09x ratio. A recipe tuned for one brand will be noticeably under- or over-salted with the other. Salt at the end, by feel.
- Baking ratios. Flour weights, leavening, and hydration are not where models are dependable. Use AI for the idea, a tested recipe for execution.
- Allergen safety. Models hallucinate ingredient sources and processing steps. Verify gluten, nut, dairy, and shellfish yourself.
A specific failure mode worth naming: the model will quietly add an ingredient you did not list (“a splash of white wine,” “a teaspoon of fish sauce”) and not flag it. Always re-read the ingredient list against your fridge before you start; otherwise you find out at step 4.
Which AI to use
For a one-off weeknight recipe, the default model on any of the big three chatbots is more than enough. As of June 2026:
| Tool | Best for | Cost | Notes |
|---|---|---|---|
| ChatGPT (GPT-5.5) | All-around recipe + photo of fridge | Free tier works; Plus $20/mo | Reads uploaded photos; “Thinking” mode for trickier substitutions |
| Gemini 3.1 Pro | ”What’s in season,” current grocery prices | Free with Google account; AI Pro $19.99/mo | Built-in live web search is the real edge for seasonal swaps |
| Claude (Sonnet 4.6) | Tight instruction-following on “use only my list” | Free tier works; Pro $20/mo | Best at not sneaking in extra ingredients when told not to |
| SuperCook | No prompt, no signup, match pantry to recipes | Free | Pick from 2,000+ ingredients or dictate them by voice |
| DishGen | Chat-based, no account needed | Free | Type ingredients, get a formatted recipe instantly |
If you do not want to write a prompt at all, SuperCook and DishGen are the two free dedicated tools worth knowing. SuperCook ranks recipes by how many ingredients you already own; DishGen generates a fresh recipe from a plain-language list. For everything else — substitutions, scaling, “make it less spicy” — a general chatbot is more flexible.
The fridge-photo shortcut
Typing your inventory is the slow part. Instead, open the fridge door and the crisper drawer, take one photo, and upload it to ChatGPT, Claude, or Gemini. All three default models read images. The model inventories what it can see, and you fill in the pantry staples it cannot (oil, soy sauce, rice in the cupboard). Then attach the prompt below.
Two cautions: vision models miss items at the back and misread half-used jars, so confirm the inventory it lists before you cook. And it cannot read an expiry date you cannot see — your nose still outranks the model on whether the chicken is fine.
What to feed the AI
- Proteins you have, with rough quantity (“2 chicken thighs, ~1 lb”)
- Vegetables you have, including the ones slightly past prime
- Pantry staples that should count as available (oil, soy sauce, garlic, common spices) — list them so the model does not pretend they are absent
- Carbs (rice, pasta, bread, tortillas)
- Total time you have, honestly — 25 minutes including chopping is very different from 25 minutes of stovetop time
- Equipment available (one pan? oven? air fryer? instant pot?) — the recipe shape changes
- Dietary constraints and dislikes, specific ones (“no cilantro” matters; “healthy” doesn’t)
- How spicy you actually like it — the model’s default heat level is usually wrong for someone
Copy-ready prompt
Generate a 1-recipe dinner from these ingredients only. Do not add ingredients I did not list.
Ingredients: [paste — proteins, vegetables, pantry, carbs]
Time available (including prep): [minutes]
Equipment: [one pan / oven / instant pot / air fryer]
Constraints: [dietary, dislikes, heat level]
Return:
1) Recipe name — descriptive, not marketing ("Soy-Glazed Chicken with Rice and Peppers," not "Weeknight Hero Bowl").
2) Total time: prep + cook, broken out. Be honest — slice / dice for someone who is not a line cook.
3) Ordered steps with each step's time estimate. Start each step with the active verb.
4) A parallel-cooking note — what's happening in the pan while rice simmers — so things finish together.
5) Salt / pepper / acid pacing: when to taste, where to add the acid, what to do if it tastes flat at the end.
6) Two "if you have it" optional upgrades (one fresh herb, one fat or acid).
7) Substitution table: 3 ingredients I might be missing and what to swap to.
8) One sentence on what could go wrong and how to recover.Shorter variant — “what do I make right now”
I have [3-5 ingredients] and [minutes] minutes. One pan only.
Give me one dish, 5 steps max, with the exact moment I should taste for salt. Skip the headnote.Sample output
A useful step block: “Step 2 (3 min): start rice in saucepan with 1.5 cups water and a pinch of salt; cover, low heat. Step 3 (5 min): while rice is going, dice onion and bell pepper, slice chicken into 1-inch strips. Step 4 (8 min): high heat, oil shimmering, sear chicken 4 min per side; add onion and pepper at the flip. Step 5 (taste check): when chicken is opaque and peppers are blistered, taste — if it’s flat, add soy sauce by the teaspoon, not salt. Step 6: serve over rice; squeeze of lemon if you have it.”
A useful substitution table: “No bell pepper → any sturdy vegetable (zucchini, broccoli, mushrooms); add 1 minute to cook time. No soy sauce → 1 tablespoon Worcestershire + a pinch of sugar. No fresh garlic → 1/4 teaspoon garlic powder, added when oil is hot, not cold (it burns when cold).”
How to refine
- Force “use only what I listed”: “Re-read the recipe and remove any ingredient I did not give you, including ‘splash of wine,’ ‘pat of butter,’ or ‘fresh herbs.’ If a substitution is needed for flavor, name it in the substitution table, not silently in the steps.”
- Realistic prep time: “Add 50% to prep time for a home cook who is also tired and chopping for the first time today. If a step says ‘mince garlic,’ that is 90 seconds, not 10.”
- Parallel-cook the steps: “Re-order so the longest-running item starts first and faster steps happen during its downtime. Rice should be on by step 2, not step 5.”
- Anchor seasoning to a taste moment: “Replace ‘season to taste’ with one specific moment: ‘after the vegetables soften but before the protein finishes — taste, then decide.’ Give me one fix if it tastes flat and one if it tastes harsh.”
- Add the recovery line: “End with one sentence on the most common failure for this dish (overcooked protein, mushy vegetables) and what to do the moment you notice.”
Common mistakes
- Trusting cook times exactly — they swing 20-40% with stove, pan, and starting temperature; always check doneness, not the clock
- Salting at the start at full amount — sauces and reductions concentrate; salt at the end, because you can add more but cannot remove it
- Asking AI for a recipe then mentally changing five things — at that point you are writing the recipe; tell the model your changes up front and let it re-plan
- Forgetting to declare pantry staples — leave out oil and salt and the model writes recipes that “need 2 tablespoons of oil” as if oil is something you buy each time
- Ignoring the “use only my list” rule — re-read the ingredient block before you start; the model will sneak in wine, butter, or fresh herbs
- Using AI for baking — flour, leavening, and hydration ratios are not reliable; use AI for ideas, a tested recipe for execution
- Skipping the taste check — the model can’t taste; that step is yours, and salt plus acid almost always need one more nudge at the end
- Trusting AI for medical dietary restrictions — gluten, nut, dairy, and shellfish allergens need ingredient-by-ingredient human verification, not “AI says it’s gluten-free”
FAQ
- Can I trust AI for dietary restrictions? For preference-level (vegetarian, dairy-free, low-carb) it is usually fine. For medical-level (celiac, severe nut or shellfish allergy), verify every ingredient yourself; AI hallucinates ingredient sources and processing steps.
- Does the fridge-photo trick actually work? Yes, on the default models of ChatGPT, Claude, and Gemini as of June 2026. Open the door and crisper drawer for one clear photo. The model misses items at the back, so confirm its inventory before you cook.
- What about baking? AI is unreliable on baking ratios — leavening, hydration, flour weights. Use it for “what could I do with bananas, oats, and yogurt” idea generation, then execute from a tested recipe.
- The recipe asks for things I do not have — what do I do? Re-prompt: “Use only the ingredients I listed. If a flavor element is missing, put the workaround in the substitution table, not the steps.” Then re-run.
- What if I have 15 minutes, not 25? Tell the model explicitly and add “no recipe step longer than 4 minutes.” The shape changes: stir-fry over noodles, not braise; sandwich, not bake.
- Will the recipe taste good? Honest answer: it will taste like food. AI recipes are reliable on technique and seasoning logic, average on creativity, and below a good cookbook on memorable flavor. For weeknight “feed me” mode, that is the right tradeoff.