alexei-led/brainstorming-ideas

Brainstorming Ideas

Follow the base skill. This Claude overlay only defines tool use and platform-specific behavior.

Claude tool rules

• Use AskUserQuestion for every choice point. Do not write numbered menus and ask the user to type 1, 2, or 3.
• Ask one question per AskUserQuestion call.
• Use single-select for one path, multi-select for multiple risks/constraints/goals, and allowOther for custom answers.
• Use TaskCreate and TaskUpdate to track phases when the session has more than two steps.
• Use Read, Grep, and Glob before asking questions that the repo can answer.
• Use Write or Edit only after the user approves the exact design note, CONTEXT entry, or ADR target.
• Use external research tools only after the user chooses research or asks for it.

Suggested interactive questions

Initial question:

• Header: Idea type
• Question: What would you like to brainstorm?
• Options: New feature, Modification, Integration, Plan grill/debate, Exploration
• Allow Other: yes

Assumptions check:

• Header: Assumptions
• Question: Which assumptions are wrong or risky?
• Options: All correct, Some wrong, Not sure, Defer this
• Allow Other: yes

Next-step checkpoint:

• Header: Next step
• Question: How should we proceed?
• Options: Explore codebase, Research solutions, Explore then research, Skip to approaches
• Allow Other: yes

Approach choice:

• Header: Approach
• Question: Which approach fits best?
• Options: Recommended option, alternative option, minimal/YAGNI option
• Allow Other: yes

Design validation:

• Header: Validate design
• Question: Does this section look right?
• Options: Yes continue, Needs changes, Go back, Stop here
• Allow Other: yes

Implementation handoff:

• Header: Next steps
• Question: What should happen next?
• Options: Create worktree, Create plan, Save design note, Done for now
• Allow Other: yes

Optional exploration

If the user chooses codebase exploration, run a bounded read-only scan. Prefer direct Read/Grep/Glob; use a subagent only for broad scans.

Prompt shape for a broad scan:

Quick scan only. Find project structure, relevant flows, conventions, integration points, and tests for: <idea>. Return 5 bullets with file paths. Do not edit.

If no relevant code appears, say no prior implementation found; do not fabricate patterns.

Optional research

If the user chooses research, use Perplexity or WebFetch with a scoped query. Summarize sourced patterns before proposing approaches. If live retrieval is unavailable, say so and continue from local context only.

Capture rules

• For CONTEXT.md entries, ask for approval of the exact one-sentence definition before editing.
• For ADRs, require all three: hard to reverse, surprising without context, and a real trade-off.
• For design notes, write only concise Problem, Chosen approach, Trade-offs, Open questions, and Testing strategy.

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.
• If the user asks for task sequencing, state that it is outside this skill's scope.
• If the user asks for generic technology research, route to researching-web.
• If the user asks for docs-only work, route to documenting-code.
• If the user stops mid-session, offer BRAINSTORM PAUSED or a short design note.

Execute this collaborative brainstorming workflow now