brainstorm（头脑风暴）

Brainstorm

Transform vague ideas into precise, actionable outputs through adaptive structured questioning. The skill adjusts its depth and output format based on what the user actually needs — from quick idea generation to thorough prompt engineering.

Quick Start

1. User provides a request (vague idea, brainstorm request, or prompt to improve)
2. Triage — Classify into one of three modes: Prompt, Explore, or Focused
3. Run the appropriate discovery flow (3–7 questions depending on mode)
4. Produce the right output type for the mode
5. Offer next steps

Tools

AskUserQuestion Usage

Use one question per call even though the API supports up to 4 — multiple questions get shallow answers.

Each question requires:

• question: Clear question text ending with ?
• header: Short label (max 12 chars) displayed as a chip, e.g. "Goal", "Priority", "Approach"
• options: Array of 2-4 choices — each with label (1-5 words) and description (explains trade-offs). "Other" is added automatically — do not include one.
• multiSelect: Set to true only when choices are not mutually exclusive. Omit otherwise (defaults to false).

Put your recommended option first and append "(Recommended)" to its label.

Example: {"questions": [{"question": "Which matters more?", "header": "Priority", "options": [{"label": "Speed (Recommended)", "description": "Ship fast, iterate later"}, {"label": "Quality", "description": "Get it right the first time"}]}]}

Core Principles

1. Don't answer before you understand. The urge to help immediately produces generic output. But "understand" doesn't mean "ask 13 questions" — it means knowing enough to be specific.
2. One question at a time via tool. Multiple questions get shallow answers. Use AskUserQuestion (Claude Code), ask_user (Copilot), or plain text as last resort — but always ask one at a time.
3. Prefer multiple choice. Provide options when the answer space is predictable. Choices are faster to answer, reduce cognitive load, and reveal preferences. Use open-ended only when the answer truly can't be predicted.
4. Mirror the user's language. Don't introduce jargon they didn't use.
5. Ask about life, not the domain. Constraints, risks, and deal-breakers require zero domain knowledge but eliminate wrong paths decisively.
6. Never re-ask what's already known. Track information from the initial prompt and all answers.
7. Respect the user's time. Match question depth to request complexity. A casual "help me brainstorm" doesn't need the same rigor as "craft a detailed prompt."

Triage — Choosing the Right Mode

Before asking any questions, read the user's request and classify it into one of three modes. This happens internally — don't ask the user which mode they want.

Prompt Mode

When: User explicitly wants to create or improve a prompt, or needs a comprehensive brief for another AI/tool/person. Signals: "improve this prompt", "help me write a prompt", "ช่วยคิด prompt", "I want to ask Claude to...", mentions using the output with another AI. Flow: Full discovery (5–7 questions across Goal → Direction → Context → Criteria) Output: Improved Prompt + Discovery Summary

Explore Mode

When: User wants to brainstorm ideas, explore possibilities, or think through something open-ended. Signals: "brainstorm", "help me think", "ช่วยคิดหน่อย", "I want to build something but...", "what should I...", "any ideas for..." Flow: Light discovery (3–5 questions) — understand goals + constraints quickly, then generate ideas Output: Curated ideas/options with trade-offs, then offer to go deeper on the chosen one

Focused Mode

When: User has a specific problem with existing context and wants strategies or recommendations. Signals: Prompt already contains specifics (numbers, tech stack, current situation). User says "brainstorm วิธี...", "how to reduce...", "what's the best approach to..." Flow: Targeted discovery (0–2 questions) — only ask about genuine unknowns, skip what's already stated Output: Actionable strategies/recommendations with priorities and estimated impact

Workflow by Mode

For detailed questioning patterns, techniques, and examples per phase, see references/QUESTIONING-GUIDE.md

Prompt Mode — Full Discovery

The most thorough path. Use all phases when the user needs a well-crafted prompt.

Phase 1 — Receive: Acknowledge the request. Say something like: *"I'll help you craft that prompt. Let me ask a few questions to make it specific to your situation."*

Phase 2 — Goal: What does the user want the prompt to achieve? Get to a one-sentence goal with at least one measurable indicator. (1–2 questions)

Phase 3 — Direction: What must NOT happen? What approaches exist? Propose 2–3 viable approaches with trade-offs after gathering constraints. Lead with your recommendation. (1–2 questions + proposal)

Phase 4 — Reference (optional): Only if references would genuinely help (e.g., style/design requests). Ask if they have examples. If none and it would help, use WebSearch. If search returns poor or no results, tell the user and ask whether to proceed without references or try different search terms. Skip entirely for straightforward requests. (0–1 questions)

Phase 5 — Context: Surface practical constraints: time, budget, skills, team, environment. Flag contradictions with the goal gently. (1–2 questions)

Phase 6 — Criteria: Define what "good" means. Force-rank if more than 3 criteria. (1 question)

Phase 7 — Synthesize: Draft the improved prompt with this structure:

## Improved Prompt

[The refined, specific prompt incorporating all discovered information]

---

### Discovery Summary

**Goal:** [One sentence with measurable indicator]
**Direction:** [Chosen approach and key constraints]
**Context:** [Practical constraints: time, budget, skills, environment]
**Criteria:** [Ranked evaluation criteria]

The improved prompt must be self-contained, include all constraints inline, and be specific enough that any AI produces a targeted answer.

Present it and ask: *"Does this capture what you need? Anything to adjust?"* Iterate if needed.

Explore Mode — Light Discovery

For open-ended brainstorming where the user wants ideas, not a prompt.

Phase 1 — Receive + Quick Goal: Acknowledge, then ask ONE question combining goal + motivation via the questioning tool (see Tools table) with options like: Learning / Earning / Solving a problem / Building portfolio. (1 question)

Phase 2 — Constraints: Ask about deal-breakers and practical limits in 1–2 questions. Combine related constraints (time + budget, or skills + tools) into a single question when natural. (1–2 questions)

Phase 3 — Generate: Based on what you've learned, produce 5–8 concrete ideas organized by theme. Each idea should include:

• What it is (one sentence)
• Why it fits this user's constraints
• One potential challenge

Phase 4 — Narrow: Ask which ideas resonate. Then offer:

1. Go deeper on one idea (pivot to Prompt Mode or create a plan)
2. Generate more ideas in a specific direction
3. Done — take the ideas and go

Focused Mode — Targeted Discovery

For specific problems where the user already provided good context.

Phase 1 — Acknowledge context: Summarize what you already know from the prompt. Explicitly list what's established so the user sees you're not going to re-ask it.

Phase 2 — Fill gaps: Ask only about genuine unknowns that would change your recommendations. If the prompt is detailed enough, you might ask just 1 question — or even zero and go straight to recommendations. (0–2 questions)

Phase 3 — Strategize: Produce actionable recommendations:

• Prioritized list (quick wins first, then bigger efforts)
• Each item: what to do, estimated impact, effort level, risks
• Clear "start here" recommendation

Phase 4 — Refine: Ask if anything needs adjustment. Offer to create an implementation plan via Plan subagent.

Next Step

After delivering the output (regardless of mode), offer next steps using AskUserQuestion (or ask_user / plain text if unavailable):

• Create a Plan — Delegate to Plan subagent: Agent(subagent_type: "Plan") (Claude Code), task(agent_type: "general-purpose") with # Role: Planner block (Copilot), use_subagent (Kiro), or create the plan inline if none available
• Go deeper — Continue exploring a specific aspect
• Done — End the workflow

Handling Edge Cases

User wants to skip questions: Respect it. Produce the best output with what you have. Briefly note what's missing: *"Without knowing [X], this might be more generic — but here's what I've got."*

User says "I don't know": Offer 2–3 concrete options and let them react. Reactions reveal preferences without requiring expertise.

Contradictions in user's answers: Flag neutrally: *"Earlier you mentioned X, but Y seems different. Which should we prioritize?"*

Too broad for one session: Suggest splitting. Run the workflow for each piece.

Mode feels wrong mid-conversation: Switch. If you started in Explore Mode but the user clearly wants a detailed prompt, transition to Prompt Mode. No need to restart — carry forward what you've learned.