The Model Is a Mirror

Whatever you bring to the prompt, the model hands back with better grammar.

[ trace // field response ]

Whatever you bring to the prompt, the model hands back with better grammar. Vague intent in, vague prose out. Half-formed plan in, confident-sounding half-formed plan out. The system does not invent rigour you did not supply; it polishes whatever shape you handed it until the edges look intentional.

This is why working with a model is a diagnostic for your own thinking. The places where the output goes soft are the places your brief went soft. The hallucinations cluster around the questions you did not bother to answer for yourself. The fix is upstream of the prompt, almost always.

Treat the conversation like a code review of your own intent. If the model keeps misunderstanding, the bug is in the spec. Sharpen the brief before you sharpen the prompt, and the prompt will mostly write itself.