SavourAI: Redesigning How AI Talks to Cooks
Chat-based AI makes a poor cooking companion. It assumes rather than asks, ignores what's in your kitchen, and gives up when it gets things wrong. SavourAI is a multimodal, context-aware cooking assistant that fixes all three.
A problem I lived through.
After moving to the UK and living independently for the first time, cooking became a daily challenge I hadn't anticipated. I relied on whatever I could find, text recipes I had to scroll through with floury hands, YouTube videos I couldn't pause at the right moment, video calls home asking my family to walk me through steps. None of these worked well together.
When I tried using ChatGPT to help, I ran into a new set of problems: it didn't know what I had in my kitchen, it defaulted to Western cooking conventions without asking, and when I challenged a wrong answer, it apologised and gave me a different wrong answer.
That experience became a design question: what would a cooking AI look like if it was actually designed for how people cook?
What I found when I tested what already exists.
Systematic experimentation with ChatGPT and Gemini using realistic cooking prompts. Each prompt was documented with screenshots and specific failure notes.
I asked for "a quick, light dinner under 30 minutes." ChatGPT defaulted to Western food norms without asking my cuisine preference. When I specified Indian food and lactose intolerance, it suggested a prawn dish under the vegetarian section, a critical factual error that immediately destroys trust.
When I typed "peppers" intending black pepper, the AI assumed chilli peppers and built a recipe around the wrong ingredient. Rather than asking "which kind of peppers?", it resolved the ambiguity silently and incorrectly. There was no recovery path, I had to restart the entire conversation.
Recipes used terms like "alternating with milk," "mix until just combined," and "cook until golden brown" with no explanation. Prep time was buried inside cooking time with no separation. Oven preheat duration was omitted entirely. None of this works for a novice cook mid-task with one free hand.
"I added the written amount of milk but still the pasta has lumps. No easy way to ask the AI for help." Once an AI output was wrong or unclear, the only option was to type a new prompt and start over. There was no structured feedback mechanism, no ability to flag an issue, no corrective flow.
"Current AI cooking tools prioritise fluent output over correct output. They assume rather than verify, and offer no path to recovery when they get it wrong."
What the research says about cooking and AI.
This project was grounded in secondary research and systematic AI experimentation, not primary interviews. The insights below come from academic literature, peer-reviewed HCI papers, and documented user behaviour patterns.
People today rely on a fragmented mix of apps, social media, video platforms and digital recipe tools, most switch between 3–4 sources during a single cooking session. Video recipes are preferred over text, they're more engaging and reduce ambiguity at critical steps. Current AI tools are used for practical guidance (49% of ChatGPT usage is "asking," seeking advice rather than task completion), but cooking is underserved because it requires sensory, iterative, real-time support that text-only interfaces cannot provide.
Research identifies five recurring failure modes: over-trust in AI suggestions without verification; miscommunication from vague inputs ("cook until done" is uninterpretable without visual context); cultural and linguistic mismatch, since AI training data skews Western and English-language; cognitive overload from long text instructions during hands-busy cooking; and loss of human-like reassurance when AI feels rigid where users need confirmation. (Sources: Califano et al., 2025; Ngo Cong-Lem et al., 2024; Moran, 2023)
From documented user behaviour: users need to know how many servings a recipe makes, what allergens are present, exact ingredient quantities in standard units, cooking time separated from prep time, difficulty level before they commit, step-by-step progress tracking with the ability to mark steps done, a way to ask questions mid-cook, and visual confirmation that they're doing each step correctly. None of these are available in standard chat-based AI tools.
Survey: cooking habits & AI expectations.
Not beginners, not confident improvisers. This shaped the target user: someone capable, but uncertain, who wants guidance without being talked down to.
Before cooking starts, not during it. SavourAI's homepage is oriented around this finding: the primary entry point is "what should I cook today," not a recipe search field.
Split almost evenly between those two brackets. Nobody is planning an elaborate meal. This set the ceiling for what the AI recommends by default, and why cook time is a filter, not a preference left for the user to figure out.
Almost always both at once. This confirmed the substitution engine and ingredient-led search aren't add-ons, they're the primary recovery behaviour.
No hedging. The appetite was there, the question was whether the product would earn it.
These five map directly to SavourAI's core feature set, every one of them validated by the data before a single screen was designed.
Two cooks, two different constraints.
"Am I doing this right?"
Living independently for the first time. Lacks confidence, unfamiliar with cooking terminology ("what is smoked paprika?", "what does sauté mean?"), relies heavily on visual guidance.
- Follow a recipe confidently without getting lost
- Understand unfamiliar terms without leaving the flow
- Feel reassured they're doing each step correctly
- Ambiguous instructions written for experienced cooks
- No visual confirmation mid-step
- No way to ask "am I doing this right?" of a text-based AI
"I already told you this, why are we starting over?"
Cooking under real constraints, limited ingredients, 30 minutes, specific dietary needs that change by day. Doesn't want to re-enter preferences every time.
- Get a relevant recipe fast, without re-explaining constraints
- Have saved preferences respected automatically
- Cook within a strict time budget
- Recipes that ignore constraints already specified
- Re-entering preferences every session
- Generic suggestions that don't match what's on hand
Both users share the same underlying need: an AI that meets them at their level of knowledge, in their kitchen, with what they actually have.
Four principles that drove every decision.
Collect constraints upfront, diet, cuisine, ingredients, time, skill level, so the AI doesn't have to assume. Assumptions are where miscommunication begins.
Sliders, toggles, image scanning and dropdowns give the AI clearer, more precise signals than open-ended prompts. Reducing reliance on free text reduces the surface area for misinterpretation.
Miscommunication is expected in human-AI interaction. The interface needs a clear, low-effort path back, structured feedback, corrective flows, ingredient substitution, without losing context or starting over.
Voice input, camera-based guidance, visual step breakdowns and minimal text all acknowledge the physical reality of cooking. Users cannot type long prompts with floury hands.
How might we respond?
How might we make the AI understand what's actually in the user's kitchen, not what it assumes?
How might we separate meal discovery from step-by-step cooking guidance, since these are different user intents?
How might we give users a structured way to recover when the AI output is wrong?
How might we make cooking guidance work hands-free and in real time?
How might we help novice cooks confirm they're doing each step correctly, without having to ask?
Working out the structure before the pixels.
Low-fidelity passes over the home screen, testing how much input to ask for up front, where filters and ingredient capture should sit, and whether inspiration belonged on the same screen as the request.
Mapping the full journey, from login to finished meal.
The IA was designed around two distinct user intents, finding a meal idea versus executing a specific recipe, with recovery flows built in throughout. The full flow was mapped before any screen was designed: login and onboarding into a homepage with two modes (Meal Ideas and Cook Now), each branching into its own generation path, converging on a step-by-step cooking flow with a live video mode, and a recovery flow reachable from any point via a flag-response menu.
The recovery flow, flag response → feedback menu → ingredient review → regenerated output, is reachable from any point in the journey, not just after a finished recipe.
Five decisions that shaped the product.
Separating meal discovery from recipe execution
Users come to a cooking app with two completely different intents, "what should I cook?" versus "help me cook this specific thing." A single chat interface treats both the same, producing irrelevant results for both.
Two primary entry points, Meal Ideas and Cook Now, separated at the top level. Each flow optimises for a different cognitive mode: exploratory browsing versus focused execution.
Separating intent at the start of the journey produces more relevant AI outputs and reduces the mental overhead of managing a conversation doing two things at once.
Structured input instead of a blank prompt
Asking a hungry user to write a detailed free-text prompt, "quick, vegetarian, Indian, under 30 mins, no dairy, low carb", before seeing any results creates friction at exactly the worst moment.
A structured AI interface with dropdowns for meal type, cook time, cuisine and diet, plus sliders for effort level, spice level and health level. Voice, camera and text remain available but are supplementary. The system builds the prompt on the user's behalf.
(Fogg Behaviour Model) Motivation is highest at the moment of decision, ability must match. Reducing the effort of input at peak motivation is what gets users to actually engage rather than abandon.
Ingredient scanning: grounding AI in reality
AI generates recipes assuming all required ingredients are available. This is the single most common source of misalignment between AI output and user context, "I don't have all the ingredients" was the most frequent complaint in the research notes.
Camera-based ingredient scanning as a first-class input method. Users photograph what they have; the AI detects and lists them for confirmation; recipes are generated from what is actually in the kitchen.
Grounding AI output in physical reality, not assumptions, is the most direct intervention against the miscommunication that produces wrong recipe suggestions. (Common Ground Theory: shared context must be established before communication can succeed.)
Miscommunication recovery flow
When an AI output is wrong, current interfaces offer no structured way to correct it. Users must retype a new prompt, losing context and starting over, or they abandon the conversation.
Every recipe card has a three-dot menu with "this response is incorrect." Selecting it opens a structured feedback screen, a short list of the most common failure reasons (wrong ingredients / allergen / dietary mismatch / too complex / something else). Based on the selection, the system enters a corrective flow: review the ingredient list, substitute missing items, confirm, regenerate.
(Error recovery, Nielsen's Heuristic #9: help users recognise, diagnose and recover from errors.) The feedback categories were designed from the most common failure patterns documented in the research, not invented. This makes recovery feel natural rather than like filing a complaint.
Live video mode during cooking
"Am I doing this right?" is the question a novice cook needs answered most frequently during cooking, and it's the question no text-based AI can answer. "Cook until golden brown" means nothing without visual confirmation.
A live video mode accessible during the step-by-step cooking flow. The user points their camera at what they're cooking; the AI observes the visual state and provides contextual guidance, "the pasta does have some lumps, keep stirring and add a little more milk."
Cooking is a sensory, visual, physical activity. Text is a fundamentally limited medium for guiding it. Real-time visual feedback closes the gap between AI instruction and user reality, the gap that generates the most frustration in current tools.
The rest of the system.
Quick entry points on the homepage, Quick (under 15 mins), Healthy, Fancy, Easy, generate a structured prompt on the user's behalf. Removes the blank-page problem for users who don't know what they want yet.
When a user is missing an ingredient mid-flow, the AI suggests alternatives that maintain the recipe's intent, pasta replaced with noodles, cauliflower instead of broccoli. Users can substitute or remove without abandoning the cooking flow.
Spice tolerance, dietary restrictions, cuisine preferences and allergens are captured once at onboarding and applied to every generation, no re-entering per prompt. Users can override for any session without losing their saved defaults.
What I'd carry into the next project.
The most important design work on this project was not the happy path, it was the recovery flow. Most AI products treat incorrect outputs as edge cases. They are not. In cooking especially, where outputs need to match physical ingredients, skill levels and cultural contexts, the AI will frequently miss. Designing a clear, low-effort, structured path back from that failure is what separates a tool users can trust from one they abandon after the first wrong answer.
Replacing free-text prompts with sliders, toggles and selectors wasn't about simplifying the interface, it was about creating a shared vocabulary between the user and the AI. When users express "spice level: medium" through a slider, the AI receives an unambiguous signal. When they type "not too spicy," the AI guesses, and guessing is where miscommunication begins. Every input mechanism in SavourAI was chosen because it reduces the surface area for the AI to make an incorrect assumption.
This project began with my own experience as a novice cook living independently for the first time. That wasn't a substitute for research, it was the starting hypothesis that gave every subsequent research finding a home. The most powerful design insight was not from an academic paper. It was from the moment I typed "peppers" and got a recipe built around chilli peppers when I meant black pepper, and had no way to tell the AI it had misunderstood me. Personal experience, treated critically, is legitimate research material.