Recitly: Recall It. Own It.

Recitly is an AI-powered learning app for iPhone that turns topics, notes, files, and photos into structured flashcard decks. The development team designed the AI generation pipeline, integrated a commercial LLM API, implemented adaptive spaced repetition, and connected the resulting content to several active-recall modes.

The app is available in the Education category and uses a freemium model. Free users receive a daily allowance of AI generations, while paid subscriptions extend access and are managed through Apple ID.

The Background

Creating useful flashcards manually takes time. Students need to identify key concepts, write concise definitions, remove duplicates, and structure the material before they can begin learning.

Traditional flashcard apps reduce the amount of paper involved but often preserve the same manual workflow. They also tend to treat every card equally, even though users remember some concepts quickly and repeatedly struggle with others.

Recitly was created to solve both problems. First, its AI flashcard generator converts raw learning material into editable term-and-definition pairs. Second, an adaptive review system changes when and how each card appears based on the learner’s previous answers.

During onboarding, users select a goal such as exam preparation, language learning, school or university study, professional certification, or personal development. This context helps shape the product experience without limiting the subjects users can add later.

The result is not simply a storage tool for digital cards. It is a learning system built around AI-assisted content preparation, active recall, and spaced repetition.

The Goal

The central technical task was to create one generation system that could process different source formats while producing a consistent flashcard structure.

Multimodal Content Pipeline

Recitly accepts four input types: a short topic prompt, pasted text, an uploaded file, or photos of learning materials. Users can add up to five images at once and combine content from all of them into one deck.

Each source enters through a separate input handler, but the extracted content is normalized before generation. Text from prompts and notes can move directly into the processing layer, while files and images require content extraction and validation. Multiple images are grouped in the correct order so that pages from one chapter or set of notes can be interpreted as a single source.

This unified pipeline keeps the downstream AI logic independent of the original format. Once the material is prepared, the same generation layer can produce 5, 10, 15, or 20 cards.

Anthropic API Integration

The generation layer uses the commercial Anthropic API. Instead of training a custom language model, the team built a controlled LLM integration for mobile apps around an existing provider.

Requests are prepared on the server rather than directly inside the iOS client. This protects API credentials, centralizes model settings, and allows the backend to validate input size, enforce quotas, and handle failed requests consistently.

The model is instructed to return structured term-and-definition pairs rather than free-form text. Prompt rules define the requested number of cards, preferred answer length, language, source boundaries, and duplicate handling.

This structure makes the output easier to validate before it reaches the mobile application. It also supports future model changes without rebuilding the learning interface.

Prompt Engineering and Quality Control

Generating a card is not only a summarization task. The system must identify information worth remembering, separate closely related concepts, and keep each answer specific enough to be assessed.

The prompt design therefore focuses on several requirements: grounding cards in the submitted material, avoiding repeated ideas, preserving the source language, and producing definitions that can work across different study modes.

After generation, users see a preview of the complete deck. They can review every term and definition, remove unsuitable cards, and save only the selected results.

This review step is an important part of the AI workflow. The interface also warns users that AI may produce inaccurate information and that important facts should be checked.

AI-Generated Test Distractors

The test mode required a separate generation approach.

A standard flashcard contains one correct term-and-definition pair. A multiple-choice question also needs several wrong answers that are believable enough to test knowledge without becoming ambiguous.

Recitly generates these distractors with dedicated prompt rules. The model receives the correct answer and relevant deck context, then creates alternatives from the same subject area. The output must remain grammatically consistent, distinct from the correct response, and plausible to someone who has not fully learned the topic.

This is a more demanding custom AI features development task than basic Q&A generation. Weak distractors make the correct answer obvious, while overly similar options may create more than one defensible response.

Spaced Repetition and Card Mastery

Once cards are generated, Recitly uses an SM-2-based scheduler to determine when they should return.

The algorithm tracks card performance through values such as repetition count, ease factor, review interval, and next review date. Correct answers move the card forward, while mistakes reduce its mastery state and return it to the review queue sooner.

The user-facing learning mode simplifies this process into two mastery levels. The first stage uses four-answer multiple choice. The second checks the concept through a true-or-false task. A wrong answer lowers the card’s level, requiring another review.

This approach connects classic spaced repetition app development with a progression system that is easy to understand inside a mobile interface.

Multiple Learning Modes

The same deck can be studied in several ways without regenerating its content.

Flashcard mode supports quick swipe-based review. Learn mode adapts tasks to the card’s mastery level. Test mode uses AI-generated distractors, while Input mode requires the learner to type the answer manually.

Users can also practise through a timed matching game or a fast true-or-false mode. Progress indicators show the current question, correct answers, and the final percentage, such as 6 correct responses out of 10.

This allows one generated dataset to support recognition, recall, typing, matching, and timed decision-making.

AI Quotas and Cost Control

AI generation has a direct operational cost, so free access needed to be useful without allowing uncontrolled API usage.

Free users receive three AI generations per day while manual flashcard creation remains unlimited. Premium removes the daily AI restriction.

The quota is enforced on the backend rather than only in the app interface. Each authenticated generation request checks the user’s plan, current usage count, and quota reset period before contacting the LLM provider.

The server records a successful generation atomically, preventing simultaneous requests from exceeding the allowance. The iOS interface displays the remaining balance, but the backend remains the source of truth.

This rate-limiting approach helps balance AI costs while allowing new users to test the main feature before subscribing.

Privacy and Data Handling

Learning materials may contain personal notes, unpublished documents, or photographed textbook pages, so the AI flow was designed with data handling in mind.

Recitly sends generation requests through Anthropic’s commercial API rather than a consumer chat account. Anthropic states that commercial API inputs and outputs are not used to train its models by default unless the customer explicitly opts in or submits material as feedback.

The app also allows users to delete their account and associated data directly from the product.

The Development Team

AI and Prompt Engineer

The AI specialist designed prompts for flashcard creation, explanations, and multiple-choice distractors. Their work also covered structured output validation, source grounding, generation limits, and handling different input formats through one pipeline.

iOS Developer

The mobile developer implemented onboarding, source selection, image and file uploads, generation states, deck previews, learning modes, progress tracking, and subscription access. They also connected the mobile interface to the backend AI services without exposing provider credentials inside the application.

UI/UX Designer

The designer translated complex learning logic into a clear mobile flow. This included source selection, card-count controls, generated deck review, learning-mode navigation, mastery feedback, result screens, and visible AI accuracy warnings.

Project Manager

The project manager coordinated product requirements across AI generation, mobile development, learning logic, privacy, and monetization. The role also involved prioritizing features and aligning model behavior with the expected user experience.

The Results

Recitly was launched as an iPhone application in the App Store’s Education category. It combines multimodal flashcard generation, an SM-2-based review system, six learning modes, progress tracking, and freemium subscriptions in one product.

Users can turn a topic, document, text passage, or group of photos into a study deck without manually rewriting every concept. They can review the generated cards before saving them and then practise the same material through different active-recall formats.

The completed product demonstrates a practical approach to AI app development: the LLM handles content transformation and distractor generation, while deterministic application logic manages validation, quotas, card mastery, scheduling, subscriptions, and user control.

This balance makes Recitly an AI-powered learning app, rather than a generic chatbot placed inside an education interface.