Sometimes building an AI bot looks like this: You spend two days hunting a “complex” bug, wipeSometimes building an AI bot looks like this: You spend two days hunting a “complex” bug, wipe
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
Tanya's avatar
Tanya Shpak
• 19h
Sometimes building an AI bot looks like this:
You spend two days hunting a “complex” bug, wipe server history, clear the Telegram chat, reset memory to zero, restart the bot, burn a few of your own pennies on tests, almost give up…
…and then you realize the whole problem was this:
The bot doesn’t see /start as a command. It sees “/start” as plain text.
So you just sit there like: Of course. Naturally. Exactly how I wanted to spend those two days.
The bot kept clinging to the old branch, and after a restart it would jump right back into the middle of the previous flow — as if nothing happened and we’ve been here forever.
Because this is a no-code stack and I don’t have backend access, I couldn’t “fix it properly” from inside the system. I had to think differently: How do I send the right signal to the model throughthe limitations of the platform?
At some point I went straight to Qwen and asked how to explain what I needed. And… on the third attempt, we finally aligned.
Now the bot actually resets mid-flow, breathes нормально, thinks, and returns to the correct branch. It’s still : bit overcautious (it double-checks where it could be bolder), but that’s the next iteration.
The deeper I go, the more I enjoy this inverted logic.
With traditional code, you expect: run a command → get a predictable action.
With LLMs, it’s flipped:
If you want more freedom, liveliness, and weirdness — you often have to loosen control. If you want it to “just behave like a normal human” — you suddenly discover the model has its own definition of normal and human.
And that’s the fun part:
An LLM is not code with a personality. And not a person without a body.
It’s a different kind of interlocutor — and sometimes you have to learn to think backwards.
In no-code chat stacks, treat “/start” as data unless the platform guarantees command handling.
promptengineeringllmnocode
Back to feed
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started