Habr AI publishes Rozitta Parser case study: how vibe coding eats sleep and time

Habr AI has published a personal case study about how an attempt to save an important Telegram chat turned into six months of building a custom application. An architect with no programming background built Rozitta Parser with the help of several neural networks, and she honestly describes how much time, energy and everyday resources this kind of vibe coding consumes.

How the project began

It started with a simple task: the author needed to export the contents of a Telegram group where old information was meant to disappear once a new stream of posts began. She asked a neural network to write a parser, even though not long before she did not know what Telethon, API keys or `pip install` were. The quick script solved the local problem, but almost immediately it began to accumulate new requirements: message filtering, file downloads, support for different chats and a more convenient way to launch it.

From there the project followed a path typical of vibe coding: one AI model wrote code, another helped structure prompts, and a third looked for alternative tools and explained why one stack was better than another. DeepSeek was used for the first versions and HTML prototypes, NotebookLM as a supporting knowledge base, Gemini for phrasing prompts, and Claude for heavier assembly and refactoring. The one-off script grew into a full desktop program with its own name and a mascot, a pink robot frog called Rozitta.

What the parser grew into

Over six months, Rozitta Parser became a multi-module Python application with a GUI, data export and local media processing. The author stresses that the project has moved far from the original idea of "saving a conversation": it is now a tool that tries to turn Telegram into more than an archive, a convenient knowledge base for further work, reading and uploading to AI services such as NotebookLM. In effect, it is already a personal pipeline connecting the messenger, a file archive and data preparation for later analysis.

• Export of conversations to DOCX, JSON, Markdown and HTML
• Download of files, images and voice messages from chats
• Local audio transcription via faster-whisper, without the cloud
• Splitting large exports into chunks for upload to AI services

The current stack looks quite mature: Python 3.11, PySide6 for the interface, Telethon for working with the MTProto API, SQLite for data storage, python-docx for export and faster-whisper for local transcription of voice messages without the cloud. The app works with SOCKS5 and MTProto proxies, and can also slice large exports into chunks of 150 thousand words to get around the limits of third-party services. In essence, it is no longer a "script for myself" but a complex product that has to be maintained like real software.

The cost, and help

The strongest part of the text is not technical but human. The author calls the taxi to work her "fatigue tax": because of late-night coding sessions she began arriving late, losing sleep, spending money on eating out and dropping out of family life. The article does not romanticize the endless build. Instead, it is a story of how creative excitement in AI development easily turns into an "either code or sleep" mode, especially when the project grows faster than your understanding of architecture, Git and the limits of your own time.

"Fatigue tax" is not a metaphor.

The turning point came when the project went onto GitHub and she asked the community for help. After a question on Habr Q&A, the app gained a live helper who took on bugs, the interface, proxies, builds and explaining basic processes such as GitHub Actions. In parallel, an interactive module dependency map and clearer documentation appeared. One of the author's main conclusions is blunt: AI speeds up the start, but over the long run it does not replace documentation, understanding of the language or the involvement of real developers.

What this means

The Rozitta Parser case shows the current limit of vibe coding well: getting into development is easier than ever, but the price of a mistake has also grown. Neural networks already help beginners build a working product, yet without structure, rest and human support, such a project quickly starts consuming not only time but the rest of life as well.