The Casino Problem: Ditching the One-Armed Bandit for Real Wins

We’ve all felt the promise of AI IDEs. Whether you're using Cursor, Windsurf, or Copiloty, the productivity boost can be amazing, but there are times it can also be frustrating. You write a prompt and press the submit button, you wait for the response, but you do not get what you wanted. So you try again, alter the prompt, press submit, and yet again, not what you wanted. It feels frustrating, and it's also akin to sitting at a one-armed bandit at a casino in Las Vegas, where each pull of the handle is the hope that the next one will be a winner.

Why does this happen? There are a huge number of reasons, and I do not want to boil the ocean here, but one significant issue is the 'model auto model selection' feature. This feature will automatically select the model to use based on the complexity of the task, the size of the task, and other factors. On the surface, the auto model selection feature seems like a good idea. And yes, it is and it has a place and time to use it. But if you're new to using AI as part of your development workflow, then you should not be using the auto model selection feature. The reason is that you will not be able to learn how the model reacts to your input. Learning to understand how the model reacts and generates is key, key to being able to craft prompts to have the model perform the task you ask it to. If the model selection happens automatically, it's nearly impossible to learn. Each time model selection happens, you may get different models for different tasks.

To become a solid AI-assisted software developer requires the ability to know how to craft input to have a model perform tasks and carry out these tasks, including code generation. And this is just not possible with the model selection process being automatic. You are in an endless game of hope. If you set the model selection process to be manual and select a solid capable model like Gemini, GPT5, or Anthropic, you can get to know intimately how the model thinks. This is achieved with time and interacting with the model. And there is no way to achieve this with using auto model selection. With the model changing, you're just confusing yourself and playing a game of chance and luck. In this blog post, we will dig into understanding how switching to manual makes a difference, and how selecting a single model makes a real difference.

The Flaw in the Promise

I'm talking about the feeling of being in a Las Vegas casino, sitting in front of a one-armed bandit. You pull the handle (click "generate"), hold your breath, and hope for a win. Sometimes you get a jackpot—a brilliant block of code that's exactly what you needed. Other times, it's a dud. The most frustrating part is you never know which one you're going to get. There is no pattern, and it feels impossible to develop a working style, as if you're just gambling on the next response. It’s a frustrating cycle where a simple task like writing a comment might trigger a full-scale model switch, introducing an unpredictable delay and an inconsistent result.

This is the central drawback of the auto model selection. It’s a desirable option in theory, but its randomness prevents us from developing the crucial muscle memory of selecting the right tool for the job. As you do not perform the model selection, you never get a feel for how a specific model is behaving, and you don’t get a complete understanding of its strengths and weaknesses. What makes things worse is that as different models get selected for other tasks, you can't tie the output to any particular model. It makes learning impossible.

The Breakthrough: Picking a Partner and Sticking with Them

The most pragmatic solution, the one that delivers real results right now, is to turn off the auto feature and make a deliberate choice. Stop gambling. Instead, we choose a single, competent model and make it our partner.

I've found great success with foundation models like Anthropic Opus or Gemini 2.5. By committing to just one of these and using it exclusively for a time, we gain something invaluable: consistency and predictability. We move from hoping the tool gets it right to knowing what it will do. This isn't a step backward; it's a focused step forward, optimizing our partnership with the AI.

This is how we find our rhythm. By dedicating ourselves to one model, we begin to learn its personality. We start to understand its nuances: what kind of prompts it responds to best, what kind of tasks it excels at, and where its limitations are. We're not just throwing prompts at a black box; we're actively learning how to communicate with it. This deliberate practice is how you develop a deep understanding of its character, how it thinks, and how to get the most out of it.

The Payoff: Real Results from Deliberate Practice

The benefits of this deliberate practice are tangible and immediate. When you stop switching models, the quality of your work becomes more consistent. You start to see a direct correlation between the prompts you write and the code you get back. Your time is no longer wasted on vague or unhelpful suggestions, because you have developed a mental model of how your partner thinks. The experience becomes less about troubleshooting the tool and more about getting the job done. It's a genuine win that makes the daily work of coding feel a lot less like an uphill battle.

This focused effort is where the real power of prompt engineering comes from. You’ll develop a feel for how the selected model is performing, its strengths and weaknesses. You'll start to fully understand how you can prompt the model to generate precisely what you need. By taking control of the model selection, you get back to building with a confident, steady pace, leaving the one-armed bandit behind.

The Future of the Auto-Pilot

Now, this isn't to say that auto-selection is a lost cause. The vision is solid: an AI that can intelligently triage tasks and apply the right model for the job. As these systems mature and as we gain more insight into their inner workings, the autopilot will become more reliable and predictable. For now, however, the most effective path forward is to be the pilot yourself. Taking the stick, choosing your model, and learning its personality is the quickest way to find success today.