The Model Is a Mirror

Whatever you bring to the prompt, the model hands back with better grammar. It does not invent rigour you did not supply.

I ran this loop rewriting mystic-bytes essays: vague brief in, confident-sounding vague essay out. The prose improved. The mechanism did not. Passages went soft exactly where my spec went soft — which layer owned invalidation, what the staleness budget was, what failure looked like on a tired 2am deploy. The model polished the shape until the edges looked intentional. The hallucinations clustered around questions I had not answered for myself.

This is why model output is a diagnostic for your own thinking, not a substitute for it. Soft output marks soft intent. Confident wrongness marks missing constraints. The fix is almost always upstream of the prompt — in the brief, the acceptance criteria, the named case you forgot to include.

Treat the conversation like a code review of your spec. If the model keeps misunderstanding, sharpen the brief before you sharpen the prompt.

— JV · Dark Heart Labs.