Video summary
AI-генератор фанфиков в Cursor за 21 день: показываю свой вайб-кодинг стек
Main summary
Key takeaways
Summary of the Video “AI-генератор фанфиков в Cursor за 21 день: показываю свой вайб-кодинг стек”
Project Overview
Oleg, an experienced server developer turned live coder, presents his new project called LMATF, a web application that generates fanfiction stories based on user input.
- Users specify a universe (e.g., Harry Potter, Cyberpunk 2077) and a plot idea.
- The app generates chapters of approximately 2000 words each, taking 5–10 minutes to read.
- The generated text quality is described as quite high.
Project Architecture
The project consists of three main parts:
- Backend: Built with FastAPI and PON.
- Frontend: Developed using Next.js with Tailwind CSS.
- Telegram Bot: Manages premium subscriptions via a paid access bot.
The stack was chosen for familiarity and good AI-assisted coding compatibility. Next.js was preferred over pure React for easier routing.
Development Process and AI Tool Usage
- The main coding assistant is Cursor, an AI code generator.
- Oleg emphasizes a structured, step-by-step approach when using Cursor:
- Fully understand the task before prompting the AI.
- Highlight relevant files to keep Cursor focused.
- Separate backend and frontend changes.
- Write down important implementation notes (e.g., database migrations).
- Features are developed semi-automatically by iterating in small increments and testing frequently.
- For complex features, tasks are broken into minimal viable products (MVPs), tested early and often to avoid unmanageable code.
Asynchronous Task Handling
- Long-running AI generation calls are handled asynchronously.
- Tasks have statuses such as pending, success, and error.
- Users can refresh and check progress without losing data.
- Oleg notes challenges with Python’s asynchronous behavior and stresses the importance of verifying async compatibility.
Neural Network Models and API Integration
- Used OpenRouter to unify multiple model APIs (Claude, OpenAI, Deepseek), enabling easier switching and testing.
- Tested various models for cost and quality:
- Naklo 3.7: Best for initial text generation.
- GM Mini 2.5 Pro: Cheaper but less accurate with Russian language.
- Strategy:
- Generate the first 5 chapters with Naklo 3.7.
- Use GM Mini for subsequent chapters to maintain style and reduce cost.
- To overcome token limits (about 60 chapters max), implemented a summary-based context feeding:
- Summaries include plot events, consequences, character motivations, and “Chekhov’s guns” (plot elements to be developed later).
- Only the last three chapters are included in full; earlier chapters are summarized to reduce token count while preserving story depth.
Additional Tools and Services Used
- MCP servers: Enhance model reasoning and consistency (Sequential Thinking server).
- Player: Tool for web navigation and data retrieval by the model.
- Break search: Google-like search to improve AI’s coding accuracy.
- Kotext: Library documentation tool.
- GPT O3: Highly intelligent model used for business advice, UI/UX suggestions, project feasibility analysis, and concept design.
- GPT-4 image generator: Creates website visuals, logos, avatars, and backgrounds based on prompts generated by GPT O3.
- Plausible Analytics: Self-hosted alternative to Google Analytics integrated for visitor and interaction tracking.
Key Insights and Conclusions
- Modern AI tools can accelerate development speed by about 10x, enabling complex projects to be built in weeks instead of months.
- Success depends on choosing the right tech stack, having experience with tools, and applying a disciplined development approach.
- The project and workflow showcased represent early but promising steps in AI-assisted coding and content generation.
- Oleg plans to continue developing new projects and sharing insights via his Telegram channel.
Main Speaker
Oleg — a developer and live coder sharing his experiences and technical insights about building an AI-powered fanfiction generator using Cursor and other AI tools.