A Method for Generating Full-Stack Mobile Systems using Large Language Models

A Method for Generating Full-Stack Mobile Systems using Large Language Models 📂 Repository Structure . ├── 1.app-description/ # Input app descriptions (*.txt) ├── 2.feature-extraction/ # Feature extraction scripts and outputs │ ├── script-extract-features.py │ ├── prompt-feature-extraction.txt │ ├── requirements.txt │ └── extracted-features/ ├── 3.app-generation/ # Generated application artifacts │ └── apps.zip # Archive of generated apps ├── 4.feature-evaluation/ # Evaluation summaries and charts │ ├── generate-feature-classification-summary.py │ ├── prompt-feature-evaluation.txt │ ├── script-evaluate-features.py │ └── evaluated-features/ │ ├── ai/ # AI-evaluated feature classifications │ ├── human/ # Human-evaluated feature classifications │ ├── ai-vs-human-comparison-summary.txt │ └── ai-vs-human-classification-chart.pdf ├── 5.code-evaluation/ # Code evaluation raw data (extracted from SonarQube Community Edition) └── README.md 📝 How to Use This Repository Prerequisites Python 3.10+ OpenRouter API key (for feature extraction) Set your key: export OPEN_ROUTER_API_KEY="your_key_here" Setup Feature extraction: cd 2.feature-extraction python -m venv .venv source .venv/bin/activate pip install -r requirements.txt Evaluation summary generation (from the repository root): pip install matplotlib Workflow 1) Prepare app descriptions Ensure your input app-description files are available in 1.app-description/ (one app per .txt file). 2) Extract features From the 2.feature-extraction/ environment: python script-extract-features.py This creates or updates extracted outputs in 2.feature-extraction/extracted-features/. 3) Generate applications Run your generation process to produce the generated apps archive in 3.app-generation/ (currently packaged as apps.zip). 4) Evaluate features Use the prompt in 4.feature-evaluation/prompt-feature-evaluation.txt to classify each extracted feature against the generated application source code. Save per-app results as: 4.feature-evaluation/evaluated-features/ai/*_ai-evaluated-features.txt 4.feature-evaluation/evaluated-features/human/*_human-evaluated-features.txt 5) Build evaluation summary and charts From the repository root: python 4.feature-evaluation/generate-feature-classification-summary.py This command generates: 4.feature-evaluation/evaluated-features/ai/global-feature-classification-summary.txt 4.feature-evaluation/evaluated-features/human/global-feature-classification-summary.txt 4.feature-evaluation/evaluated-features/ai/ai-feature-classification-chart.pdf 4.feature-evaluation/evaluated-features/human/human-feature-classification-chart.pdf 4.feature-evaluation/evaluated-features/ai-vs-human-comparison-summary.txt 4.feature-evaluation/evaluated-features/ai-vs-human-classification-chart.pdf 📱 Generated apps The applications used in the experiment were selected to cover a diverse set of domains, ensuring variability in functionality and complexity. The dataset spans 17 categories, including Books, Business, Design, Education, Entertainment, Finance, Food & Drink, Health & Fitness, Lifestyle, Medical, Music, Navigation, News, Photography, Productivity, Shopping, Social, Sports, Travel, Utilities, and Weather. These categories encompass a wide range of interaction patterns, such as content consumption (e.g., Entertainment and News), user-generated content (e.g., Social and Photography), transactional services (e.g., Finance and Shopping), and domain-specific applications (e.g., Medical and Health & Fitness). Application Category Description Attempts (back | front) Close2BobbieGoods Books Digital coloring book app offering interactive drawing and creative activities. 2|2 Close2Duolingo Education Language learning platform with gamified lessons and exercises. 1|1 Close2Threads Social Social media platform for sharing short text-based updates. 1|1