deep-grill

name	deep-grill
description	Stress-tests plan decisions node-by-node with focused questions. Walks each decision point surfacing assumptions and weak spots. Use when the user asks to deep grill or stress-test a plan.
allowed-tools	Read, Glob, Grep, AskUserQuestion, TodoWrite
disable-model-invocation	true

Deep Grill

Walk each branch of a plan's design tree, asking focused questions at every decision node and exploring the codebase to resolve them.

Table of Contents

• Step 0 — Create progress todos
• Step 1 — Resolve the plan file
• Step 2 — Read and build design tree
• Step 3 — Walk design-tree branches
• Step 4 — Present deep grill summary

Instructions

Step 0 — Create progress todos

Before doing any other work, use TodoWrite to create todos for each step of this session. This gives the user visibility into progress and ensures no step is skipped.

Create the following todos (all starting with status: "pending"):

• Step 1: Resolve plan file
• Step 2: Read and build design tree
• Step 3: Walk design-tree branches (deep grill session)
• Step 4: Present deep grill summary

Mark each todo status: "completed" as you finish that step.

Step 1 — Resolve the plan file

Use the first match from this priority order:

1. Argument: If a file path appears in $ARGUMENTS or the user's message, treat it as the plan file path and read it directly.
2. Currently open file: If no path was given, check whether a .md file is currently open or selected in the IDE (provided via context). If it exists and its content looks like a plan (contains headings like ## Implementation, ## Plan, ## Steps, ## Context, or ## Summary), use it.
3. Settings plansDirectory override: Read the "plansDirectory" key following Claude Code's settings precedence — project-local .claude/settings.local.json, then project .claude/settings.json, then global ~/.claude/settings.json. The first that sets it wins (a more specific layer takes precedence). Glob *.md files from that path and use the most recently modified file.
4. Default fallback: Use Glob on ${PWD}/docs/plans/*.md, sort by modification time, and select the most recently modified file.

If no file can be found via any of these methods, tell the user and stop.

The deep grill is designed for implementation plans. If the resolved file is a SKILL.md (filename is SKILL.md or frontmatter contains name: and description: without plan-style headings), inform the user that the deep grill targets plans, not skill files, and stop.

Announce the file: "Deep grilling plan: path/to/plan.md"

Step 2 — Read and build design tree

Read the resolved plan file and extract the design tree:

1. Identify decision nodes: Scan the plan for architectural choices, technology selections, API designs, data models, integration points, file structure decisions, and any other points where an alternative approach exists.

2. Organize into branches: Group decision nodes by plan section or concern area. Each branch represents a related cluster of decisions.

3. Present the outline: Show a brief summary of the branches found:

   Found N decision branches:
   1. [Branch name] — [brief description]
   2. [Branch name] — [brief description]
   ...
   

4. Ask the user via AskUserQuestion whether to grill all branches or select specific ones:

   • Question: "Grill all branches or pick specific ones?"
   • Options: "Grill all branches" / "Let me pick specific branches"

   If the user picks specific branches, present the list with numbers and let them choose.

Step 3 — Walk design-tree branches

For each selected branch, one at a time:

• Announce the current branch: "Branch N/M: [branch name]"
• For each decision node in the branch, ask a focused question and provide your recommended answer before waiting for the user's response.
• If the question can be answered by exploring the codebase, use Glob, Grep, or Read first, then present your finding as the recommended answer.
• After each response, check whether sub-questions exist for that branch. If so, resolve them before moving to the next branch.
• Collect all decisions and insights as you go.

After completing each branch (except the last), use AskUserQuestion to ask:

• Question: "Branch N complete. Continue to the next branch?"
• Options (first = recommended): "Yes, continue" / "No, stop and summarize"

If the user stops early, proceed directly to Step 4 with the branches examined so far.

Step 4 — Present deep grill summary

Output a structured summary:

## Deep Grill Summary

### Plan

[path/to/plan.md]

### Branches Examined

[N/M branches completed]

### Key Decisions Resolved

- [Branch]: [Decision confirmed] — [rationale or codebase evidence]

### Open Questions Remaining

- [Any branches or sub-questions the user deferred or skipped]

### Recommendations

- [Amendments to the plan based on deep grill findings]