Agent Skills: Brainstorming Ideas

Brainstorming Ideas

Turn a vague idea or draft plan into a well-formed design before coding. Keep the session collaborative, question-driven, and small enough to change direction.

Core rules

• Ask one question at a time.
• Use an interactive question tool when available; do not emulate menus in plain text.
• Inspect code before asking when code can answer.
• Offer 2-3 options with trade-offs and mark one recommendation.
• Always allow a free-text or Other answer when options may not fit.
• Cut speculative features and keep implementation task breakdown out of scope.

Interactive questions

When a runtime question tool is available, use it for every choice point:

• single-select for one path, approach, or confirmation
• multi-select for multiple goals, risks, constraints, or audiences
• free text for problem statements, plan details, or custom answers
• options plus Other when you can suggest likely answers but need flexibility

Do not ask the user to type 1, 2, or 3 unless no interactive tool is available. If no tool exists, use concise labeled options and include Other.

Load domain context

Before design questions, look for relevant project docs:

• CONTEXT.md
• CONTEXT-MAP.md
• docs/adr/
• nearest */CONTEXT.md or */docs/adr/

Read them when present. Use those terms in questions and designs. If no docs exist, create them only with user approval and only when a real term or decision is resolved.

Understand the idea

If the user did not supply a topic or plan, ask what they want to explore. Then narrow one question at a time until you can state the problem in one sentence. Prefer questions about:

• pain or trigger
• user or actor
• existing feature it builds on or replaces
• explicit non-goals
• strongest constraint

Grill or debate mode

Use only for a bounded plan, trade-off, or assumption. If none is clear, ask for one. Read references/grill-protocol.md and follow it.

Stay focused on design quality and assumptions, not implementation task breakdown. If the user asks for task sequencing, state that it is outside this skill's scope.

Surface requirements and assumptions

Use 5WH, skipping what is already clear:

1. WHO uses it?
2. WHY is it needed?
3. WHAT is the core capability?
4. WHERE does it live?
5. HOW should it work, only when a hard constraint exists?

State assumptions explicitly and ask which are wrong or risky. Use multi-select when several assumptions can be wrong.

Explore context before solutions

Find similar modules, flows, conventions, integration points, test patterns, and architecture constraints. Summarize in 3-5 bullets. Cite key paths when code shaped the recommendation.

Research external solutions only if the user asks or selects that path. If the request is only a generic technology comparison or best-practice survey, use researching-web instead.

Propose and validate approaches

Present 2-3 approaches. For each:

• What: one-sentence approach
• Trade-offs: complexity vs flexibility
• Best when: scenario where it wins

Mark one recommendation. Ask which fits best with the interactive question tool.

Then detail the chosen design in short sections, confirming after each:

1. Architecture overview
2. Data flow
3. API or interface
4. Error handling
5. Testing strategy

Apply YAGNI at each section. Remove pieces that do not solve the stated problem.

Capture outcome

If the outcome is more than a short answer, offer to write a concise design note:

docs/plans/YYYY-MM-DD-<topic>-design.md

Include only Problem, Chosen approach, Trade-offs, Open questions, and Testing strategy.

If a domain term crystallized, propose a CONTEXT.md entry and write it only with user approval:

Term:
One-sentence definition.
Avoid: overloaded synonym

Offer an ADR only for decisions that are hard to reverse, surprising without context, and came from a real trade-off.

Output contracts

Completed:

BRAINSTORM COMPLETE
Topic: <topic>
Approach chosen: <name or none>
Design note: <path or none>
Key decisions: <bullets>
Domain docs: <updates or none>
Open questions: <bullets or none>

Paused or routed:

BRAINSTORM PAUSED | ROUTED TO <skill>
Topic: <topic>
Current state: <one sentence>
Resolved: <bullets>
Needed next: <question, artifact, or target skill>

Failure handling

• Idea conflicts with domain docs: quote the conflicting terms and resolve with the user before designing.
• A constraint blocks every approach: state the blocker, what would unblock it, and ask what to relax.
• Idea is too vague: stay in understanding mode; ask one narrowing question at a time.
• No bounded plan exists for grill/debate: ask for one; do not invent opposing positions.
• User stops mid-session: offer the paused output or a short design note.