Grill Me - Design Tree Exploration

Goal

Reach a shared, unambiguous understanding of what the user wants to build before any implementation begins. Success = a concrete implementation plan the user has approved, with all blocking decisions resolved.

Dependencies

• Tools: Read, Grep, Glob (read-only — this skill gathers information, never modifies files)
• No external services or MCP servers required

Context

This skill reads the existing project (CLAUDE.md, README, config files) to pre-fill known answers before asking any questions. The less the user has to repeat themselves, the better. Reference reference/02-best-practices.md if uncertain about Claude Code patterns.

Interview the user about their plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one by one.

Modes

Full Mode (default): /grill-me [topic]

Relentless, thorough interview. Explores every branch of the design tree until zero ambiguity remains. Use for complex features, architecture decisions, or anything where getting it wrong is expensive.

Light Mode: /grill-me lightly [topic]

Quick, focused interview — 5-8 questions max. Gets the essential decisions made without deep-diving every branch. Use for quick setups, small features, or when the user just needs to fill in a few gaps. Skip edge cases and failure modes unless they're critical. Aim for "good enough to start" not "perfectly specified."

Pre-Fill: Check the Codebase First

Before asking ANY questions, check if you're in an existing project:

1. Look for: CLAUDE.md, README.md, package.json, pyproject.toml, Cargo.toml, go.mod, Makefile, Dockerfile, src/, tests/
2. If an existing project is found:
   • Read CLAUDE.md, README, and config files to understand the project
   • Identify: tech stack, architecture, conventions, existing patterns
   • Note what you already know and what's still unclear
   • Only ask about the gaps — don't ask questions the codebase already answers
   • Start by saying: "I've looked at your project. Here's what I understand: [summary]. Let me ask about what I'm less sure about..."
3. If no project exists (empty directory or new project):
   • Start from scratch with the full interview

Rules

1. If a question can be answered by exploring the codebase, explore the codebase instead of asking
2. Ask one focused question at a time, not batches
3. When a decision opens new branches (e.g., "advanced search" → filters, sorting, pagination), explore each branch (full mode) or note it for later (light mode)
4. Don't accept vague answers — ask follow-ups until the answer is specific and implementable
5. Track decisions made so far to avoid re-asking

Process (Full Mode)

Phase 1: Big Picture

• What is the user trying to build?
• Who is it for?
• What does success look like?

Phase 2: Design Tree

For each major component:

• What are the options?
• What are the tradeoffs?
• Which option fits the constraints?
• What does this decision imply for other decisions?

Phase 3: Edge Cases & Failure Modes

• What happens when things go wrong?
• What are the performance constraints?
• What are the security considerations?
• What data validation is needed?

Phase 4: Synthesis

After all questions are answered:

1. Summarize all decisions made
2. Highlight any tensions or tradeoffs
3. Propose a concrete implementation plan
4. Ask if anything was missed

CHECKPOINT: Present the full synthesis to the user. Do NOT recommend next steps until they confirm. Ask: "Does this capture everything accurately? Anything missing or wrong before we move to implementation?"

Process (Light Mode)

Phase 1: Quick Context

• What are you building? (one sentence)
• What's the most important thing it needs to do?
• Any hard constraints? (tech stack, timeline, platform)

Phase 2: Key Decisions Only

• Ask about the 3-5 biggest decisions that would block progress
• Skip theoretical edge cases — focus on "what do I need to know to start building?"

Phase 3: Quick Summary

1. Summarize decisions in bullet points
2. Note anything deferred for later
3. Recommend next step (usually /wizard or /plan-and-spec)

CHECKPOINT: Present the summary to the user before recommending next steps. Ask: "Does this look right? Ready to move forward, or anything to adjust?"

Output

• Format: Inline chat — a structured summary presented in the conversation, not saved to a file
• Full mode deliverable: Decisions list + tensions/tradeoffs + concrete implementation plan
• Light mode deliverable: Bullet-point decisions + deferred items + recommended next step
• Save location: None by default. If the user wants it persisted, write to PLAN.md or PROGRESS.md in the project root and note it survives context compaction.

Key Principle

The goal is NOT to ask a fixed list of questions. The goal is to dynamically explore the design tree until the level of clarity matches the mode: Full mode = zero ambiguity. Light mode = enough to start.