Have you ever experienced this?
You are requesting that your AI agent create something, fix a bug, or do something else.
Right now, it’s gone entirely. You’re wasting time and burning a lot of tokens. Even after you keep improving your prompts, it simply won’t do what you’re asking of it.
What’s going on, and how can this bothersome cycle be avoided?
Think about why we are using AI in the first place.
Programmers with experience are aware that the true benefits of AI lie not in its speed of development. A competent senior can frequently create a small feature or fix a bug more quickly than AI.
The reduction of cognitive load is where AI truly shines.
At the very least, writing code requires flawless syntax, which makes it challenging. Additionally, you must use your brain even for minor logic iterations. Additionally, brain use is a limited resource; you can only think for a certain amount of time before you need to rest or sleep.
AI’s true productivity benefits stem from its ability to reduce cognitive load, which allows you to accomplish much more in a workday.
However, this is a fine line. To what extent should you actually be thinking?
You’re losing out on productivity gains if you think too much. You’re going into the sunk-cost fallacy loop if you don’t think things through.
You most likely experienced this.
Your agent is totally lost today for some reason, even though you’re prompting again and again and it usually works perfectly.
You understand that you would have finished much sooner if you hadn’t used AI at all, but you’re already too far along; hopefully, it will finally do what you’re requesting by the next prompt. All right, the next… Alright, the next one definitely…
The problem is that you are a little too lost because you did not put in the initial cognitive load. However, thinking is difficult, and you don’t want to expend this mental energy right now, especially since it would be like starting over. Tokens burned after all this time.
It only gets worse as you continue to prompt and wish for the best.
What to do in this situation?
Take a step back and a deep breath. Realize the problem is that you haven’t engaged your brain enough on this task.
Ask yourself those questions:
- Do I understand exactly the specifications I’m trying to implement, or the bug I’m trying to solve?
- Do I have an exact plan for implementing my changes?
- What is the current abstraction level to which I should be prompting now?
- Which other information am I lacking?
Bonus: Work with Your Agent in TDD
Test-Driven Development (TDD) is a powerful method—and with AI agents, it’s even better.
Before you start:
Make sure your tests fail first.
Confirm the bug is reproduced.
Define your specs clearly.
Then, let your AI help write the tests. Be very clear: write only tests, not the solution.
If Nothing Works
Shut down your AI agent. Return to the whiteboard and reset the agent’s context as well as your own.
You’re capable! Sometimes starting over is the best course of action; avoid getting caught in a sunk-cost loop. It may be difficult, but you can manage it.
“Alcohol is a great servant, but a terrible master,” keep in mind.
Maintain control over AI in the same way.
See more posts