There’s a troubling pattern in how some people use AI. They set up vague prompts and hope for the best. When it doesn’t work, they tweak and try again. Tweak. Try. Repeat until something lands.
There’s a name for this: prompt roulette.
What Prompt Roulette Actually Is
Prompt roulette is iterating on prompts with no real definition of success in mind. Without knowing what success looks like, you judge by feel. The outcome becomes random.
It’s treating AI like a slot machine. Pull the lever. Hope for a win. It’s addictive, expensive, and frustrating. And it’s becoming normalized. Companies are rewarding it.
The Tokenmaxxing Trap
Successful companies are bragging about their AI spend. They’re mistaking token usage as a measure of AI adoption success.
More AI usage equals more productivity, right? No. Token spend is input, not output.
Using tokens as a success metric doesn’t mean you’re moving faster. It just means you’re consuming more. And it actively encourages prompt roulette. The more you iterate, the more tokens you burn, the “better” you look.
We see this constantly. Our CEO hears from other companies about their AI budgets, their token consumption. And he asks: why aren’t we spending more?
Here’s the answer: our metric is what ships, not how many tokens we burned. We spend less than $500 per person per month on AI and we move fast. That’s not because we’re cheap. It’s because we’re intentional.
Prompt Roulette vs. Disciplined AI Usage
Vague prompt, hope for the best
No definition of success. Judged by feel. Outcome is random. More tokens burned, more iterations, more cost - no guarantee of a better result.
Precise prompt, defined criteria
Success is defined before the prompt is written. Output is reviewed against criteria, not by feel. Predictable results, maintainable code, lower token spend.
How Disciplined AI Usage Actually Works
We don’t hand off a spec to AI and walk away.
We follow the same process we always have. Break the feature into user stories. Define acceptance criteria. Understand the existing codebase. Create an implementation plan. Only then do we ask AI to help execute.
But here’s what’s different: the developers using AI do the thinking. They read the context. They understand what success looks like. They write precise prompts because they’ve done the work to know what they’re asking for.
Then they review the output against the criteria they set. They validate it. They judge it against a standard, not by feel.
This requires more upfront work. More thinking. More clarity. But it produces something that prompt roulette never will: predictable results and maintainable code.
The real work isn’t pulling the lever. It’s knowing what you’re asking for before you ask.
The Difference It Makes
Prompt roulette might feel faster in the moment. But it’s expensive and actually takes longer.
The winners aren’t burning more tokens. They’re using their judgment and skill to make better use of them. They do the hard work first: defining what they need, understanding the context, setting success criteria. Then they use AI to execute faster.
It’s not revolutionary. It’s just discipline applied to a new tool.
The Question
So here’s what I want to ask you:
Are you worried about how your dev team is using AI? Are they playing roulette, or are they being intentional about it?
Let’s talk.