Use when a request is vague, a beginner is using AI like a search box, or first answers are weak because the ask was underspecified, and when the user says "help me ask this", "improve my prompt", or "why are my answers bad". Diagnoses the raw ask against six levers (role, goal, context, constraints, format, definition-of-good), asks 2-4 high-leverage clarifying questions, routes any sensitive context through prompt-safe-send before it is attached, writes 2-5 explicit gradeable success criteria handed forward to ai-result-eval, and emits a one-line teach-back so the user learns the pattern instead of depending on it.
