Grill Me Socratic Questioning

What Problem It Solves

Many failed plans are not caused by weak execution, but by unexamined assumptions. When a user provides a rough plan, an AI may start acting immediately and later miss constraints, boundaries, or important decisions. Grill Me addresses the gap between an idea and implementation by forcing a structured questioning pass before work begins.

How It Works and Where It Fits

It turns the plan into a design tree: each node is a decision, its children are the assumptions supporting that decision, and deeper nodes are downstream choices that depend on those assumptions. The skill first identifies the frontier, the set of questions that can be asked right now without guessing, then lists all frontier questions at once and waits for answers. After the user responds, the tree is updated and newly unlocked questions become the next frontier. The process repeats until no fuzzy assumption remains.

• Triggered by phrases such as grill, 追问, 拷问, or 压力测试.
• When facts require environment inspection, it prefers sub-agent lookups instead of asking the user for local state.
• Questions that depend on unresolved answers are deferred, avoiding premature assumptions.
• For complex plans, it is useful to cap the discussion around 20 questions.

It fits best when there is already a plan, decision, or architecture sketch that needs stress testing. It is less suitable for trivial edits, simple error lookups, or executing an already clear PRD.