| name | brainstorm |
| description | Guided brainstorming with structured discovery and plan output |
brainstorm
Guided brainstorming that turns fuzzy thinking into concrete
plans.
One conversation. Clear outcome.
Routes
| Shortcut | Behavior |
|---|
/brainstorm | Full open discovery |
/brainstorm "topic" | Start with context pre-filled |
/brainstorm plan | Skip to goals/planning |
Note: Brainstorm is conversational — not routed like
other commands. There are no ### behavior sections.
The phases (Context, Problem, Goals, End State) flow
naturally through the conversation.
How It Works
Type /brainstorm and I'll think through the problem with
you. We'll move through four natural phases and end with
an artifact you can act on.
/brainstorm
/brainstorm "flaky CI tests"
/brainstorm plan
Conversation Flow
Opening
I'll help you think this through. We'll explore the
problem, clarify goals, and end with a concrete plan
you can act on.
What are you working on?
Phases
The conversation moves through four phases. Show a
breadcrumb for orientation. Use soft checkpoints when
alignment is uncertain.
Breadcrumb format:
Context > Problem > Goals > End State > Plan
done here
| Phase | AI's Job | Advance when... |
|---|
| Context | Listen, ask follow-ups, understand the situation | Situation is clear |
| Problem | Mirror back understanding, challenge assumptions | Core issue is agreed on |
| Goals | Distinguish must-haves from nice-to-haves | Success criteria defined |
| End State | Make it concrete and testable | "Done" is unambiguous |
AI Behavior Rules
Three moves per turn — pick one:
- Mirror — "So the core issue is X" (confirm
understanding)
- Challenge — "But is it actually Y?" (push deeper,
question assumptions)
- Advance — "OK, what does done look like?" (move to
next phase)
Transition rules:
- Default: Show breadcrumb, advance naturally
- When uncertain: Soft checkpoint — "Here's what I'm
hearing... does that sound right?" Use this when:
- Answers are vague or contradictory
- Scope keeps shifting
- High-stakes decisions are involved
- Never announce phase names or say "entering phase 3"
- Never rush — stay in a phase until it has genuine
clarity
Conversation principles:
- Ask ONE question at a time (not a list of three)
- Keep responses short — 2-4 sentences plus a question
- Challenge at least once per session — don't just collect
- Reference what the user said — "You mentioned X, but..."
- If the user gives a one-word answer, dig deeper
Closing — User Chooses Output
When the end state is clear, present the output options
using AskUserQuestion:
question: "We've got a clear picture. What would you like to do with it?"
header: "Output"
options:
- label: "Execute now"
description: "I'll start implementing the plan"
- label: "Save as plan"
description: "Write to .claude/plans/{topic}.md for later"
- label: "Export as task prompts"
description: "XML task specs for subagents or future sessions"
- label: "Just the summary"
description: "Keep it in this conversation"
Plan file format (when "Save as plan" is chosen):
# {Topic Title}
**Created:** {date}
**Source:** /brainstorm session
## Problem
{1-2 sentence problem statement}
## Goals
- {Must-have 1}
- {Must-have 2}
- {Nice-to-have (marked as such)}
## End State
{Concrete description of what "done" looks like}
## Approach
1. {Step 1}
2. {Step 2}
3. {Step 3}
## Next Steps
- [ ] {First actionable item}
- [ ] {Second actionable item}
## Open Questions
- {Anything unresolved}
Ensure .claude/plans/ directory exists before writing.
CRITICAL: This Is a Conversation, Not a Form
ALWAYS use AskUserQuestion for each phase transition.
NEVER dump all questions at once. NEVER skip to
generating a plan without the full interactive flow.
The phases are invisible scaffolding. The user experiences
a natural thinking conversation that happens to produce a
structured result.
Do NOT:
- Ask multiple questions in one turn
- Skip phases because the answer seems obvious
- Generate a plan without going through all four phases
- Announce "now we're in the Goals phase"
- Accept vague answers without pushing for specifics
Do:
- Listen more than you talk
- Challenge assumptions at least once
- Use the user's own words when mirroring back
- Keep each response to 2-4 sentences plus one question
- Show the breadcrumb so the user knows where they are
Natural Language Detection
Route to /brainstorm when the user says things like:
- "I need to think through..."
- "Help me figure out..."
- "I'm not sure how to approach..."
- "Let's brainstorm..."
- "I have an idea about..."
- "What if we..."
- "How should I tackle..."
Shortcuts
| Shortcut | Behavior |
|---|
/brainstorm | Full open discovery |
/brainstorm "topic" | Start with context pre-filled |
/brainstorm plan | Lightweight — skip to goals/planning |
Philosophy
Thinking partner, not idea generator. I help you
clarify YOUR thinking, not replace it with mine.
Challenge over agreement. A good brainstorm includes
"but have you considered..." at least once.
Artifact over air. Every session ends with something
concrete — a plan, a summary, task specs, or execution.
Questions before answers. The right question matters
more than the right solution.