Why one-shot queries get stuck

The common trap

Many people use an LLM like a search box that talks back: ask once, get an answer, then patch it with follow-up questions.
That works for tiny requests, but it breaks once the task hides constraints you never wrote down.

Why it breaks

• Your assumptions stay in your head instead of in the prompt.
• You care about a usable deliverable while the model only sees “generate some text.”
• You want clarification first, but the model is rewarded for answering immediately.

The upgrade

Turn a one-shot query into a minimal brief:

• What is the goal?
• What context matters?
• What must not be guessed?
• What should the output look like?
• What counts as done?

Takeaway

Stage 0 is not about fancy prompt tricks. It is about moving hidden requirements from your head into the message.