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

Tool	Purpose
ask_user	Ask the user ONE question at a time. Use choices array for multiple choice when possible.
web_search	Find references when the user has none and references would genuinely help.
task	Delegate to a Plan agent when the user wants an implementation plan. Use agent_type: "general-purpose" with plan instructions injected in the prompt.

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 ask_user. Multiple questions get shallow answers. Never embed questions in plain text — always use the tool.
3. Prefer multiple choice. Provide a choices array 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 (2–4 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 web_search. 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. Example: "What draws you to this — learning, earning, solving a personal problem, or something else?" (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 the task tool.

---

Next Step

After delivering the output (regardless of mode), offer next steps using ask_user:

• Create a Plan — Delegate to a Plan agent via the task tool with agent_type: "general-purpose", injecting plan-creation instructions and passing the full output as context
• Go deeper — Continue exploring a specific aspect
• Done — End the workflow

Plan Delegation Template

When the user chooses "Create a Plan", spawn a general-purpose agent with plan instructions:

task(
  description: "Create implementation plan",
  agent_type: "general-purpose",
  prompt: """
# Role: Implementation Planner

You are an **Implementation Planner**. Create a detailed, actionable implementation plan based on the brainstorm output below.

## Instructions
- Break down into concrete, ordered steps
- Include dependencies between steps
- Estimate effort level for each step (small/medium/large)
- Identify risks and mitigation strategies
- Provide a clear "start here" first action

## Brainstorm Output
<full output from the brainstorm session — improved prompt, discovery summary, or strategy recommendations>
"""
)

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.