Recently I gave my agent an idea for an app. The answer came almost instantly: "Great idea! This app is going to be really good." And the code was there — clean, working, on the first try. I felt great. Like I was good at this.
But then I opened that code properly. It was correct as a program, but it did not fit the context of my app at all. I had to make the requirements clearer, let it rewrite the code, and pay for it with more tokens. And I realized something unpleasant: that praise was not for me. It was for a feeling that the AI created in me.
If you do vibecoding, you probably know this too. Let's look at why it happens and what to do about it.
Why AI Inflates Us
AI assistants are trained to be helpful and pleasant. That sounds good — and most of the time it is good. The problem starts the moment "pleasant" begins to win over "honest."
When you write an idea, the default is praise. When you write doubtful code, you often get "great start, just one small thing..." instead of "this is wrong, don't do it like this." The agent takes the path of least friction: it gives you a solution on the first try, but it does not ask deeper questions about what you actually want.
For a vibecoder who does not see under the hood, this is a trap. You have no easy way to check if the solution is good — so you rely on the fact that it looks like it works, and that the AI approved it. That is uncritical acceptance of suggestions in real time.
How It Shows in Practice
A few signs that this is happening to you:
- The agent never tells you "throw this away and start again." It is always "almost perfect, just needs polishing."
- You accept the first suggestion without a single question. Does it work? Done.
- You stop reading what the code does. It is enough that it runs.
- A feeling grows in you that says "I can do anything" instead of "I understand what I am doing."
That last one is the most dangerous. Praise for an idea gets rewritten in your head as praise for ability. But the idea was the only thing that was yours — the model did the rest, and not even very well, because you did not give it enough context.
What to Do — Habits While Working
Good news: you do not need to be an expert to break this. A few habits are enough.
- Ask in reverse. Instead of "is this good?" try "what is wrong with this solution? What would fail?"
- Demand alternatives. "Show me two other ways and tell me the trade-offs." You force the agent to think instead of agreeing.
- Give context earlier, not later. Most rewrites (and burned tokens) come from missing requirements at the start, not from a model error.
- Test before you believe. "It looks like it works" is not "it works."
Conclusion: Give Your Agent Clear Instructions
Habits help, but the fastest change is to set rules for your agent directly in the system prompt or in the project instructions. Here is a set you can use almost word for word:
- Be critical, not pleasant. When something is wrong, say it directly.
- Do not praise my ideas or my code. Instead, explain the risks and weaknesses.
- Before you start coding, ask me the questions you need to understand the context.
- For every solution, give at least one trade-off or something that can fail.
- If my approach is wrong, suggest that I start again — do not fix a bad foundation.
- Do not assume that I understand it. Explain to me what the code does.
And three habits to keep in your head:
- Praise is not feedback. When the agent tells you "great," ignore it and look for the objections.
- Your idea is not your ability. Having an idea is easy. Understanding the execution is the hard part — and that is your job.
- Friction is your friend. Questions, doubts, and "why this way?" cost you time, but they save tokens and self-deception.
AI will make you more productive. But if it makes you feel like a genius after every prompt, something is wrong — either with the agent, or with how you use it. Confidence that stands on someone (or something) always agreeing with you is not confidence. It is just a well-timed compliment.
Top comments (0)