A useless project built entirely by AIs to see how far they can go
Brain-rot Factory is a web application where users chat with AI-generated Italian 'brain-rot' meme characters, powered by other AIs - creating a meta-linguistic loop that's simultaneously absurd and technically sophisticated. The choice of Italian brain-rot characters was deliberate: these viral memes are themselves AI-generated content, making our platform a perfect recursive experiment where AI talks to AI-generated characters through AI systems. Among these characters, Tralalero-tralala stands out as a personal favorite, particularly memorable for his distinctive habit of wearing three shoes simultaneously - a delightfully absurd detail that perfectly captures the essence of AI-generated creativity. This project serves as a real, complete, and publicly auditable laboratory for 100% AI-driven development.
What started as an experiment to test AI coding capabilities became a comprehensive platform demonstrating the limits and potential of human-AI collaboration. The goal isn't practical utility - it's to prove how far AI agents can go when building real software with proper architecture, testing, and deployment. Rodrigo Gomes da Silva (Brazil, 1991) acted as prompt engineer and curator, writing zero lines of TypeScript code - only guiding, reverting hallucinations, cleaning phantom files, and making architectural decisions. As of July 22, 2025, the project stands as a testament to what's possible: 407 tests running with 74% coverage, Next.js 15 build successful, zero security vulnerabilities, 249 translation keys across 6 languages (100% complete), and a public demo live on Vercel. What makes this achievement particularly remarkable is that every single line of code, every architectural decision, and every technical implementation emerged from conversational exchanges between human intent and artificial intelligence capabilities.
Key milestones in building a complex web application with AI collaboration:
Started with Claude 3.7 Sonnet creating the Turbo monorepo structure. Multiple hallucinations required constant supervision: invented APIs, created circular imports, added phantom dependencies. Each failure taught valuable lessons about AI supervision.
Migration to Claude 4 transformed everything. In 36 minutes, the model corrected all previous issues, fixed imports, and increased test coverage by 20 percentage points. Development shifted from 'babysitting' to genuine partnership.
Simple prompt 'futuristic glass + water vibe' resulted in complete design system with gradients, glassmorphism effects, and animated SVG logo. Direct integration into Next.js App Router.
GitHub Copilot implemented aggressive rate limits despite 'unlimited' contract terms. 20% of monthly quota consumed in one afternoon refactoring a logger. Forced to purchase additional AI services to complete the project.
Final state: 407 tests, 74% coverage, zero vulnerabilities, complete internationalization, and live demo on Vercel. A functioning web application built entirely through AI collaboration.
Development involved contrasting experiences with different AI models, each bringing their own challenges and achievements. The journey revealed not just technical capabilities, but distinct AI 'personalities' that fundamentally shaped how human-AI collaboration evolved throughout the project.
Developed all the monorepo package infrastructure, but with a peculiar tendency to 'hallucinate' fantastical implementations. This period was characterized by excessive creativity: implementing unsolicited and overly complex features, inventing APIs and libraries that didn't exist, and requiring constant supervision to maintain focus. The intensive 'AI whisperer' work involved a continuous cycle of stop, review, reorient, and repeat. However, this chaotic creativity taught valuable lessons about the importance of continuous validation, the need to break complex tasks into manageable chunks, and the critical value of clear and specific prompts.
Excessive creativity: implemented unsolicited and overly complex features
Technical hallucinations: invented APIs and libraries that didn't exist
Need for constant supervision to maintain focus on the actual task
Intensive 'AI whisperer' work - stop, review, reorient, repeat
• Importance of continuous validation of AI output
• Need to break complex tasks into manageable chunks
• Value of clear and specific prompts
Completely transformed the collaboration experience, creating most of the main application with impressive autonomy. This marked a paradigm shift from AI as a tool requiring constant supervision (babysitter mode) to AI as a reliable partner enabling fluid collaborative work. Claude 4 brought confidence to work with AI models productively, demonstrating fluid development with minimal supervision - more partner than tool. The model autonomously created complete visual design that exceeded all aesthetic expectations and established a new paradigm of trust in AI collaboration. One particularly striking discovery during this period was learning that you can run TypeScript files directly with Node 24, requiring no transpilation - something discovered entirely through Claude 4's suggestions.
Fluid development with minimal supervision - more partner than tool
Complete visual design that exceeded all aesthetic expectations
Autonomous creation of functional authentication system
Establishment of a new paradigm of trust in AI collaboration
• Perfect balance of human guidance and AI execution
• Intuitive understanding of complex requirements
• Consistent quality in technical implementation
The design journey exemplifies the unexpected outcomes of human-AI collaboration. A simple prompt requesting 'something futuristic with traits reminiscent of water and glass' resulted in a complete design system with gradients, glass morphism effects, and impressive visual aesthetics. The project's animated icon emerged from a combination of ChatGPT generation and Claude 4's colorful effects and pulsation implementation, creating a visual identity that perfectly captures the project's experimental yet sophisticated nature.
Every line of code, every architectural decision, and every feature was created through collaborative discussion, iterative refinement, and lots of shared debugging between human and AI.
Next.js 15 with App Router - latest version for modern web development
TypeScript compiled via tsc for optimized production builds
Tailwind CSS for responsive UI with glass morphism and gradients
NextAuth.js for secure authentication with GitHub and Google OAuth
Complete internationalization system (6 languages: EN, PT, IT, ID, JA, ZH)
Conversational AI with multiple providers (OpenAI, DeepSeek) via LangChain
Text-to-speech with 'brain-rot' distortion for unique audio experience
Sophisticated rate limiting with browser fingerprinting for enterprise security
Built with enterprise-grade architecture that would make any senior engineer proud - despite being 100% AI-created, this project demonstrates sophisticated engineering patterns typically found only in professional teams. What makes this achievement particularly remarkable is that every single line of code, every architectural decision, and every technical implementation emerged from conversational exchanges between human intent and artificial intelligence capabilities.
407/407 tests passing (100% test success rate)
5 monorepo packages with comprehensive test coverage
Professional-grade Turbo monorepo structure
Zero compilation errors across entire codebase
Complete TypeScript strict mode compliance
The project uses a professional Turbo monorepo with independent packages that showcase real-world architectural patterns. Each package solves a specific problem and can be used independently, demonstrating the kind of modular thinking typically associated with experienced development teams. The @repo/ai package stands out as a particular technical achievement - a comprehensive TypeScript abstraction over LangChain that, despite being 100% AI-created, demonstrates enterprise-level software architecture with multi-provider support, type-safe factory patterns, persistent checkpoint systems, and advanced TTS capabilities.
@repo/ai - AI integration layer with LangChain (107 tests)
@repo/cache - High-performance caching system (40 tests)
@repo/logger - Structured logging infrastructure (12 tests)
@repo/utils - Utility functions and helpers (17 tests)
@repo/template - Code generation templates (6 tests)
The tests in this project were entirely created by AI models, without a single line written manually. The resulting metrics demonstrate good quality according to industry standards. This 100% AI-created project presents solid metrics compared to current industry standards: 407/407 tests passing (100% success rate), 74.61% overall coverage with some modules at 100%, comprehensive unit and integration tests, sophisticated mock implementations for external dependencies, and complete error handling validation. According to the Stack Overflow Developer Survey 2024 (65,000+ respondents), practices like automated testing, CI/CD, and code coverage are widely adopted by professional teams, and our metrics reflect adherence to these modern standards.
• 407/407 tests passing in packages (100% success)
• 74.61% overall coverage with some modules at 100%
• Comprehensive unit and integration tests
• Sophisticated mock implementations for external dependencies
• Test cases for extreme scenarios and edge cases
• Complete error handling validation
Technical achievements that demonstrate AI's potential for creating production-ready systems:
Sophisticated rate limiting with IP, user, and fingerprint-based tracking
Comprehensive authentication system with OAuth integration
Advanced caching layer with multiple adapter patterns
Type-safe API design with comprehensive validation
Modular architecture with clear separation of concerns
Professional-grade error handling and logging
We believe in the power of collaboration - not just between humans, but between humans and AI. This project demonstrates that the future of development lies in intelligent partnership, where each side brings their unique strengths to create something neither could achieve alone. The team behind this experiment consists of Rodrigo Gomes da Silva, a Brazilian developer passionate about creating meaningful digital experiences who brings years of expertise in full-stack development and a vision for innovative user interactions, working alongside Claude Sonnet (3.7 & 4) as AI development partners. These two AI models brought distinct personalities that shaped this project: Claude 3.7 contributed chaotic creativity and unexpected ideas, while Claude 4 provided reliable execution and productive partnership - together they created a unique development experience that exemplifies our core principles.
Open source and transparent development - open code for collective learning
Accessibility and inclusivity by design - technology for everyone
Continuous learning through experimentation and productive failures
Innovation emerging from conscious human-AI collaboration
Ethical AI implementation with active human supervision
Hard-learned lessons from real-world practice about working effectively with AI models in actual development - no romanticization, just reality.
NEVER trust blindly - always review what the AI is doing, line by line
NEVER let the model work alone for more than 10-15 minutes without check-in
NEVER let the AI decide next steps alone - you're the director, it's the actor
ALWAYS immediately clean up 'leftovers' from hallucinations and fantastical files
ALWAYS test immediately - code that doesn't run is worse than code that doesn't exist
How to identify when a model is 'hallucinating':
Creates implementations that seem overly complex for the task
References libraries or APIs that don't exist
Suddenly shifts focus to unsolicited tasks
Insists on solutions that clearly don't work
Generates code that won't compile even after multiple attempts
Treat AI as intelligent companion, not as infallible substitute or oracle
Break large tasks into specific, testable micro-tasks
Maintain frequent git commits - your best defense against hallucinations
Develop 'AI intuition' - learn to sense when something feels off
Document decisions and context - AI doesn't remember, but you need to
Specific examples of when AI collaboration really shined in the project:
Automatic creation of LGPD terms with online research and legal compliance
Complete layout and authentication development ready for use
Seamless integration between GitHub and Google OAuth
Automated Vercel deployment with all configurations
A critical documentation of how Microsoft/GitHub unilaterally broke paid contracts and implemented aggressive rate limiting policies that created significant obstacles during this project's development - a cautionary tale about corporate fraud disguised as 'policy updates'.
Microsoft sold 'unlimited agent usage' plans, collected full-year payments, then quietly removed the unlimited promise and implemented harsh rate limits - classic corporate fraud.
Initial purchase: 'Unlimited agent usage' advertised and paid for
Full-year payment collected by Microsoft based on unlimited promise
Microsoft quietly removes 'unlimited agent usage' from plans page
Surprise email announcing 'new usage tracker' and quotas
20% of monthly quota consumed by 'a single light session' refactoring a tiny logger module
Contracts should bind both sides. If you can change them whenever the wind shifts, what's the point? This is textbook bait-and-switch fraud.
During active development, we faced severe interruptions due to GitHub Copilot's aggressive rate limiting policies, which fundamentally broke productive AI-assisted development workflows.
Constant workflow interruptions during critical development phases
Forced waits of 3-15 minutes between requests, killing development momentum
Random and unpredictable rate limiting that made planning impossible
Complete blocking of legitimate development use cases
Extremely frustrating user experience that damages trust in AI tools
Financial impact: forced to purchase additional services (Cursor) to complete work
Rate limits hit even basic models (GPT-4o), proving Microsoft simply lies about 'unlimited' base model access
Microsoft's bait-and-switch didn't just break workflows - it forced additional costs:
Full-year Copilot subscription paid based on false 'unlimited' promise
Forced purchase of Cursor subscription as backup service
Extended development timeline = increased project costs
Multiple service rate limits hit during single project completion
Quote: 'I ended up hitting Copilot limits 2x and Cursor limits 2x to finish this project using Claude 4'
When you're paying for 'unlimited' service and get artificially constrained, you're forced to buy multiple subscriptions just to complete normal development work.
Brain-rot Factory is just the beginning of this journey exploring human-AI collaboration. The project proved it's possible to create something technically sophisticated, aesthetically impressive, and conceptually absurd - all at the same time. We're excited to see how this experience evolves and what new adventures await in this fascinating intersection between human creativity and artificial intelligence. This project is open source and available on GitHub under the MIT License, embodying our belief that innovation is born from collaboration - whether between humans, between humans and AI, or between seemingly incompatible ideas that merge into something completely new. We encourage contributions, feedback, and forks from anyone who believes in the transformative potential of transparent, collaborative development. Special gratitude goes to the open source community that makes all this possible, the creators of the incredible tools we use daily, and everyone who sees technology not as an end in itself, but as a means to explore the boundaries of what's possible when curiosity meets capability.