Hi everyone! 👋
This is my first post here. I'm writing this not for senior developers, but for anyone who wants to gain hands-on coding experience with the help of AI tools.
Here is how my journey went
A few months ago, I had zero programming experience and had never built a Telegram bot before. As an experiment, I started "vibe coding." Today, I have a fully working Telegram bot with thousands of lines of code, modular file structure, time selection and scheduling, data encryption, a PostgreSQL database, Stripe/Paddle payments, 6 languages, and automated chart generation.
I want to share my experience: how I used multiple different LLMs in free tiers, how I bypassed rate limits, and how I escaped dead ends when AIs hallucinated, forgot context, and broke my bot.
Where It All Started
Like everyone else, I started with ChatGPT, then tried Grok, Mistral, Gemini, Qwen, and DeepSeek. I wanted to test what each model could do and for how long. I quickly ran into free-tier token limits. Also, over long distances, models tend to lose context, mess up file structures, and output broken code.
When my project was around 200–300 lines of code, working with AI was easy. But once it passed 1,000 lines, problems began. I couldn't paste the whole codebase into a free chat window anymore so the AI could see everything and find bugs.
Another issue: AIs usually forget the chat history after a few hours or start hallucinating wildly. I had to kill the chat, start a fresh one, and re-feed it the general task, snippets of code, and context.
I changed my approach and stopped overloading a single AI with everything. Instead, I split tasks between models:
Claude — The Architect: It has the best logic and understanding of overall code structure. I used it for brainstorming, system design, modular breakdown, and creating step-by-step instructions for other AIs. Also acts as a "rescuer" when other models get stuck in loops. Downside: Very strict free limits.
Qwen — The Workhorse: My main model for long distances. Given a detailed plan from Claude, Qwen sticks to it reliably. You can code with it all day long in the same chat. It holds up for about 7 days before starting to hallucinate.
Gemini — The Fresh Eyes: A secondary tool for double-checking specific code snippets or exploring new features.
Project Architecture
At the beginning, Claude walked me through setting up my local environment in VS Code:
Language & Framework: Python 3.11+, aiogram 3.x.
Database: Started with Google Sheets, migrated to PostgreSQL on Supabase (Lesson learned: use PostgreSQL from day one!).
Background Tasks: APScheduler for notifications.
Graphics: matplotlib for reports.
Payments: Paddle SDK integration.
Localization: JSON-based multi-language system (6 languages).
The Development Process & Escaping Dead Ends
For every new feature, I had Coud write a detailed step-by-step prompt/plan for Qwen. I coded in VS Code, using browser chats for the AI:
Modular structure: I kept the main file under 2,000 lines and split everything else into 200–700 line modules (handlers, database, notifications, buttons, localization).
Version control (manual): Before any change, I backed up my working files into a separate folder so I could easily rollback if things broke.
Beating dead ends: When Qwen started looping and breaking working code, I fed the problem logs and chat snippets to Claude. Claude usually spotted the logic error on the first try. In extreme cases, I "pitted models against each other," passing a broken snippet between Claude, Gemini, and Qwen until someone found the bug.
Takeaways
After a few months of part-time work, I built a working Telegram bot with thousands of lines of code from scratch.
What I learned:
AI replaces a team for simple tasks, but you must understand the project structure. Even without coding skills, you learn to think systematically and architecturally.
Use different models. Don't rely on just one. The "architect + coder + fresh eyes" stack works best.
Small steps. Write a function -> test it. If it breaks -> rollback to the last working version.
I'm happy to answer any questions in the comments!
Top comments (0)