Toque: Turn Saved Recipes Into Home-Cooked Meals

• 

  Recipes Library

• 

  Recipe Detail Page

• 

  Cooking Mode and AI Chef Assistant

• 

  Add Recipe/Ingredient

• 

  Pantry and shopping list

Inspiration

The idea for Toque came from a frustration I know firsthand — and one that Eitan Bernath described perfectly: "People save recipes constantly, but rarely cook them."

I had over 300 saved recipes across TikTok, Instagram, and YouTube. I'd scroll, get inspired, hit save, and move on. But when dinner time came, I'd open Uber Eats instead. Not because I didn't want to cook, but because the gap between a saved video and an actual meal felt impossibly wide. What ingredients do I need? Do I already have some of them? How do I follow a recipe from a 45-second video while my hands are covered in sauce?

When I saw Eitan's brief, I knew this was the problem to solve. Not another recipe database — the internet has plenty of those. What's missing is the bridge: a tool that takes the inspiration you already have and turns it into a home-cooked meal with as little friction as possible.

What it does

Toque turns saved social media recipes into home-cooked meals in three steps:

1. Save → Extract: Share any recipe link — TikTok, Instagram, YouTube, or any website — and AI instantly extracts the full recipe: organized ingredient groups, step-by-step instructions with timing, nutritional info, and difficulty level. A 60-second video becomes a structured, cookable recipe in seconds.

2. Extract → Shop: One tap adds every ingredient to a smart grocery list, automatically categorized and grouped. No more rewatching videos to write down ingredients.

3. Shop → Cook: Cooking Mode transforms each recipe into a guided, step-by-step experience with built-in timers, screen-awake lock, and an AI Recipe Assistant that answers real-time questions — tailored to the specific recipe and what's already in your pantry.

Beyond the core flow:

• AI Chef Creator generates complete recipes from your pantry ingredients, matching your time, difficulty, and dietary preferences.
• Smart Pantry Management lets you add ingredients by voice, photo, or barcode scan.
• Cookbooks organize recipes into collections by cuisine, meal type, or occasion.
• iOS Share Extension queues multiple links for batch extraction.
• Home Screen Widget shows your current shopping list at a glance.
• PDF Export shares recipes and grocery lists as beautifully formatted documents.

How we built it

Toque is built with Flutter for iOS, using a local-first architecture that prioritizes speed and offline capability.

Architecture: The app follows a clean service-layer pattern — UI Layer (Flutter widgets by feature), State Layer (Riverpod AsyncNotifier providers), Service Layer (singleton services for AI, media, storage, monetization), and Data Layer (Hive local NoSQL + SharedPreferences).

AI Engine: The core innovation is a multi-platform extraction pipeline. When a user shares a URL, the app detects the platform and routes it through the appropriate extractor — HTML cleaning for web pages, youtube_explode_dart for YouTube transcripts, Supadata API for TikTok/Instagram/Facebook video transcription. All paths converge on a unified Gemini 2.5 Flash service via dartantic_ai with strict JSON schema enforcement, ensuring consistent structured output regardless of source.

The same AI engine powers voice pantry input (audio → structured ingredients), photo pantry scanning (image → ingredient list), Chef Creator (pantry + preferences → complete recipe), and Recipe Assistant (context-aware cooking Q&A).

Monetization: RevenueCat SDK handles subscriptions with a credit-based freemium model — 5 AI operations per week for free users, with progressive soft-to-hard paywalls. Monthly ($9.99) and annual ($29.99, 3-day trial) plans unlock unlimited access.

iOS Integration: Share Extension (batch URL queuing via App Group), Home Screen Widget (SwiftUI shopping list synced via App Group UserDefaults), and local notifications for trial reminders.

Challenges we ran into

AI reliability across platforms. A TikTok video, a YouTube tutorial, an Instagram reel, and a food blog all present recipe information in completely different ways. Some have transcripts, some don't. Some have structured data, most don't. The solution was the unified AI service with schema enforcement — dartantic_ai's structured output forces Gemini to produce consistently formatted recipes regardless of input, with graceful fallback chains (human captions → auto-captions → description).

Balancing monetization with user experience. Too aggressive, and users bounce before seeing the magic. Too generous, and there's no reason to subscribe. The weekly credit system was the breakthrough: 5 operations per week is enough to form a habit but not enough for a power user. The progressive paywall — soft when credits remain, hard when depleted — converts at the moment of highest intent, not an arbitrary gate.

Social media content extraction. TikTok and Instagram don't offer public transcript APIs. We integrated the Supadata API for video transcription with a polling mechanism for async job processing and intelligent fallback to metadata-only extraction when transcripts are unavailable.

Accomplishments that we're proud of

• Universal recipe extraction: Any URL — TikTok, Instagram, YouTube, food blogs, Pinterest — gets turned into a structured, cookable recipe with a single share action. This is the feature that makes users say "wait, it actually works?"

• The AI extraction pipeline: A single unified service with schema enforcement that handles web pages, video transcripts, images, and audio — all producing the same consistent output format. One schema, one parser, every platform.

• Voice and photo pantry input: Speak your ingredients or snap a photo of your fridge, and AI structures everything — names, quantities, units, categories. It feels like magic and makes pantry management effortless.

• Cooking Mode: A full-screen, step-by-step cooking experience with timers, wakelock, and an AI assistant one tap away. It transforms the cooking experience from "following a video" to "being guided through a meal."

• The credit system design: Finding the sweet spot where free users experience enough value to form a habit, but power users feel the pull to subscribe. The progressive soft-to-hard paywall creates natural conversion moments.

What we learned

1. Schema-enforced AI output is a game-changer. Using dartantic_ai with JSON schemas turned unreliable AI text generation into a dependable structured data pipeline. This single pattern enabled the entire multi-platform extraction system.

2. The best monetization is invisible until it's not. Users should feel the value of premium before they're ever asked to pay. The credit system lets them experience the full product, making the eventual paywall feel like a natural next step rather than a barrier.

3. Building for an existing audience changes everything. When you know exactly who your user is — their age, habits, pain points, peak activity time — every design decision becomes clearer. Eitan's brief didn't just define a problem; it defined a person. Building for a real person is infinitely easier than building for an abstract market segment.

4. Fallback chains are essential for AI reliability. No single extraction method works 100% of the time. Building intelligent fallbacks — human captions to auto-generated, transcripts to descriptions, native transcription to automated — is what makes the system feel reliable to users even when individual components fail.

What's next for Toque: Turn Saved Recipes Into Home-Cooked Meals

The vision is clear: Toque becomes the default layer between recipe inspiration and meal execution. Every recipe you discover — on any platform, in any format — flows through Toque on its way to your kitchen. The MVP proves the core loop works. What's next is making it indispensable.

Built With

• claudecode
• dart
• flutter
• gemini
• revenuecat
• sharedpreferences