Just saw another prompt with "Don't make mistakes!"
But what does that actually achieve?
It forces compliance instead of understanding. A model told not to make mistakes will try harder to sound right. It won't optimize for problem-solving but for sycophancy (prioritizing user alignment over truthfulness).
We have access to the most advanced cognitive technology in history, and we're still writing if-statements like it's 1995. Just in Markdown.
Look at most CLAUDE.md files. Same pattern every time: goals, strict rules, edge cases stacked on top of each other. It works, until it doesn't. And when it breaks, it breaks exactly where the constraints failed to anticipate. The fix? More rules. Prompt rot.
Constraints can't cover what genuine context already knows.
The better investment: give the model provenance. Not just what to do — but what this is, who it's for, why it matters. General language models are general for a reason. You have to define the ground they operate on.
Stop coding your prompts. Start anchoring them.
Provenance → direction → constraints.
ps:
Mistakes are how intent gets defined. If your model never does anything wrong, you're not refining anything, you're just accepting defaults. Which kinda defeats the whole purpose of you being in the loop in the first place.
Top comments (0)