It started with a dare. A colleague claimed that 'real engineers' need at least two weeks to ship a Telegram bot, and that everyone else spends months in Jira. I had 72 hours and a free account. Three days later, I had a working sentiment-analysis bot, a spreadsheet full of real feedback from our beta users, and a total bill of just over $2.
The phrase of the season is 'vibe coding' — a workflow Andrej Karpathy described in early 2025 where you tell an AI in plain English what to build and let it write the code. FutureX sits exactly in that space: a platform that takes a conversational specification and generates a deployable workflow with connectors to Telegram, Google Sheets, and OpenAI. This article is a transparent recap of my three-day build: what I spent, where I wasted time, and what I would do differently.
The Bot: Feedback Mirror
The product was simple: a Telegram bot that accepts a forwarded message, classifies it as a bug, feature request, or 'praise/service complaint', extracts the product dimension, and appends a row to Google Sheets. The summary is just a second API interaction that generates an optional one-line recommendation. FutureX handled the orchestration, and I ran the bot on Cloudflare Workers behind Telegram's webhook.
One note for anyone trying the same stack: 'no-code' does not mean 'no thinking'. The service that made the biggest difference was Telegram's Bot API, and it is worth reading the official documentation before you build. ASI Biont supports connecting to Telegram via API — more at asibiont.com/courses.
Cost Breakdown
Here is exactly what my invoice looked like, rounded to two decimal places:
| Line item | Cost | Notes |
|---|---|---|
| FutureX | $0 | I used the 14-day trial; a paid plan starts at a monthly fee |
| OpenAI API | $2.38 | ~1,200 calls to gpt-4o-mini, including retries |
| Cloudflare Workers | $0 | Free tier includes 100,000 requests per day |
| Telegram | $0 | Bot API is free and has no per-message fees |
| Total | $2.38 | plus one large coffee |
The biggest cost driver was not tokens but retries. On day two, naive code called the model twice per message because of a timeout. After adding a retry policy, the API bill fell by 60%. If I had kept the bot in production without optimization, the monthly cost would have been roughly $8–$12. With caching and a tighter prompt, it is now under $3 at current usage.
5 Lessons from the 72-Hour Build
1. The requirement is the architecture
The first prompt produced a bot that called the API on every keystroke and sent the HTTP response directly to Telegram — needless to say, it broke. Rewriting the prompt as a structured specification (input format, output schema, error behavior) reduced token usage by about 70%. This confirms what OpenAI's documentation recommends about structured workflows.
2. Rate limits are still real
Telegram Bot API limits how fast a bot can send messages: at most one message per second per chat. My generated code ignored this and started dropping messages. Adding a simple queue with exponential backoff fixed it. The official Telegram Bot API docs list the exact limits — always verify those, not the LLM.
3. Validate every LLM output
At one point the bot returned a string like '0.9999' for a sentiment score, and Google Sheets stored it as text. I added schema validation with Pydantic, and the failure rate went from about 15% to 0.2%. This is the biggest cost-saving habit: good validation means you don't pay for a retry later.
4. Caching beats cheaper models
After several hundred messages, I noticed large parts of the system prompt were repeated in every call. Prompt caching — documented on Anthropic — reduced the effective input tokens by nearly half. OpenAI offers a similar feature; it took one line in the configuration. Do not skip this.
5. Use the 20% rule
The last six hours went into adding a database, a search box, and a dashboard. The previous 66 hours built the bot that actually worked. The version without the database processed 100% of messages and looked fine on a phone. Overengineering is a form of procrastination.
Would I Do It Again?
Yes, and I already did. The 'vibe coding' workflow is not a replacement for system design; it is a replacement for boilerplate. FutureX removed the connector glue and the hosting setup, while the hard parts — grammar, rate limits, data validation — stayed exactly where they belong: in the developer's checklist.
If you are planning to build an AI-powered bot for Telegram, do not start with a monorepo. Start with a free trial, a clear prompt, and a spreadsheet. The cost of making a mistake is now measured in cents, not sprints.
Conclusion
Three days and $2.38 taught me more than a month of planning. The bot is still running today, and it has classified over a thousand feedback messages. The lesson for product teams is simple: in 2026, the bottleneck is no longer code — it is deciding what to ask for. FutureX made the first version embarrassingly easy, and that is exactly the point.
The numbers in this article come from my own experiment. Costs may differ based on your workload, your model choices, and your patience with retries. But the overall conclusion remains: building a bot in three days is now not just possible — it is boring. And boring is good.
Comments