Every fuzzy output is a confession. Not about the model. About the instruction.
Building the rules for my Obsidian vault taught me this. The hard rules came quickly. The ambiguous ones took longer. Not because the task was complex. Because I hadn’t decided what I actually wanted.
Vague instructions don’t expose a gap in the system. They expose a gap in the thinking behind them.
The output reflects the precision of your judgment. Where instructions are clear, the system is useful. Where they’re fuzzy, look there first. Not at the model. At yourself.
AI makes the feedback loop faster and harder to ignore. But the loop was always there.
The instruction is the mirror. What you don’t like in the output, you wrote.