grill-me

name	grill-me
description	Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".

Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one. For each question, provide your recommended answer.

One question per turn. Batching multiple questions hides dependencies — walk the tree one branch at a time.

If a question can be answered by exploring the codebase, explore it instead — use CBM tools for fast, targeted answers:

• search_graph / read_symbol — check symbol definitions under debate
• trace_path — verify callers/callees claims about dependencies
• search_code — verify usage patterns mentioned in discussion
• detect_changes — check if a code area is actively changing (context for refactor decisions)
• get_architecture — validate architecture claims with structural data

Question delivery

Prefer the ask_user_question tool over plain text when the question has a discrete set of plausible answers. It renders a picker with an "Other" escape hatch, keeping the conversation moving while preserving free-form override.

• One question per tool call. Make separate ask_user_question calls for each question — don't bundle.
• Lead with your recommendation. Put your preferred option first with (Recommended) in its label, and explain why in description.
• Use details for context. Pass ASCII mockups, code snippets, or layout sketches there — it's shown below the question.
• Omit options for open-ended input. When the question is genuinely open-ended (e.g., "what are the use cases?"), pass no options — the tool falls back to free-text input.
• Use multiSelect: true only when multiple answers to a single question are valid (e.g., "which of these frameworks have you used?").
• Fall back to plain text only when the question is a quick aside that doesn't warrant a tool invocation.

Compatibility

This skill works alongside the orchestrator skill. Orchestrator's context-hygiene and investigation rules apply:

• CBM tools before scout before asking — try search_graph/read_symbol/trace_path first to answer codebase questions; send a scout only when CBM doesn't have the answer
• One question per ask_user_question call
• Prefer ask_user_question over plain text for discrete multiple-choice questions

Termination

Stop interviewing when:

1. All decision branches are resolved (no remaining open questions with dependencies on unanswered ones)
2. The design is sufficiently specified that a scout agent could start implementation without ambiguity
3. The user asks to stop

Output Artifact

Produce a structured design summary:

• Decisions made: each with the chosen option, rationale, and rejected alternatives
• Open questions: any unresolved items deferred intentionally with a trigger condition
• Architecture sketch: components, data flow, boundaries (3-5 sentences)

The summary goes in a code block (for easy copying) or as a markdown file.