When I Stopped Blaming the Model

The Blame Cycle

For months I blamed the model. When the output was wrong, I said the model was broken. When the code had bugs, I said the AI was no good. When the text was vague, I said the technology was not ready. I told myself a story: AI is overhyped, these models cannot do real work, and anyone who says otherwise is trying to sell you something.

That story kept me from seeing something I did not want to see.

The Mirror

The truth was that the model did exactly what I told it to do. When I said "write me a function," it wrote a function. Not the one I wanted, because I never said what I wanted. I named a category of thing and expected the model to read my mind.

When I said "fix this bug," the model fixed something. Not the right thing, because I never described the bug. I never said what the code should do. I never gave any context. I waved at a general area and expected a perfect answer.

Every time I got a bad result, I could trace it back to something I left out of the prompt. Every time. Not most of the time. Every single time.

The Shift

When I accepted that, everything changed. I stopped asking "why is this model so bad?" I started asking "what did I leave out?" The first question goes nowhere. The second question always has an answer. And that answer is something I can fix.

I left out the role. I left out the constraints. I left out the context. I left out the phase. I left out the format. Sometimes I left out all of them in one prompt. Then I was surprised when the output was garbage.

The model is a signal processor. It works on what I give it. When I give it a complete signal, I get a complete result. When I give it noise, I get noise back. The processor is not broken. My signal was.

The New Rule

Now my rule is simple. Before I question the output, I question the input. I read my own prompt like a reviewer. I ask one question: if I gave this prompt to a human worker with no extra context, could they give me what I want? If the answer is no, I rewrite the prompt. If the answer is yes and the output is still wrong, then I think about whether the model has a real limit.

In practice, I almost never get to that second step. The prompt is almost always the problem. That means the prompt is almost always the fix.