brain-storm

name	brain-storm
description	Brainstorms grounded feature ideas from the current codebase and saves them as wireframed notes in `brain-storm/` — the ideation step before writing specs. Use when the user asks to brainstorm ideas, generate feature ideas, explore improvements, review future opportunities, or clean up outdated brainstorm notes.
allowed-tools	Read, Write, Glob, Grep, Bash, AskUserQuestion
context	fork
agent	general-purpose

Brain Storm

Ideation before /writing-specs. Scan the codebase, propose grounded ideas, attach ASCII wireframes, save selected ideas to brain-storm/.

Modes

• Brainstorm — generate, compare, and save new ideas.
• Cleanup — detect already-implemented notes and remove them after confirmation. Triggered by "clean up brainstorm notes" or "remove ideas that are already implemented".

Hard Rules

• Ideas live in brain-storm/{name}.md or brain-storm/{subdir}/{name}.md. One subdirectory level max; filenames lowercase-with-hyphens.
• Create subdirectories only when the user asks for grouping or 5+ ideas share a domain.
• One idea = one file, and every saved idea includes an ASCII wireframe or system sketch.
• Ground every idea in the actual codebase — explore before proposing, diverge before converging.
• Never delete a brainstorm note without evidence-based verification and explicit user approval.

Brainstorm Mode

Phase 1 — Scan and ideate

1. Glob key directories (src/**, app/**, pages/**, components/**, lib/**); read package.json, README.md, and obvious entry points.
2. Glob brain-storm/**/*.md to inspect existing notes.
3. Propose 3-5 ideas, each with Title, 1-2 sentence Description, Complexity (Low/Medium/High), and a 3-5 line ASCII sketch. Make them actionable and varied in scope.
4. Ask the user which ideas to save.

Phase 2 — Deduplicate and write

For each selected idea:

• Grep brain-storm/**/*.md for 3-5 key nouns from the title/description. On a duplicate or near-duplicate, ask whether to update the existing file, create new, or skip.
• Read templates/idea-template.md and templates/wireframe-guide.md.
• Fill every section — Summary (1-2 sentences), Motivation (2-4 sentences referencing specific codebase areas), Proposed Approach (3-7 bullets, no code snippets), Wireframe (fenced ASCII, at least 10 lines), Complexity (with one-sentence reason), Open Questions (1-3 bullets).
• Write the file under brain-storm/.

Phase 3 — Report

Fill templates/report-template.md and include it in the final response. The report is part of the response, never a separate file.

Cleanup Mode

Phase 1 — Detect implemented ideas

1. Glob brain-storm/**/*.md to list all notes.
2. For each: read it, extract title and summary, then extract 3-5 implementation-indicator keywords (component names, function names, routes, API endpoints).
3. Grep those keywords in source code, excluding brain-storm/, specs/, node_modules/, .git/.
4. Mark implemented only when multiple matches clearly represent the described functionality — not TODOs or naming coincidence.

Phase 2 — Confirm and delete

• Present implemented candidates with supporting evidence.
• Ask the user — delete all, delete selected, or cancel.
• Delete only after explicit confirmation.
• Report status and evidence for each reviewed idea (Implementation Check section of templates/report-template.md).

Non-interactive Defaults

If AskUserQuestion is unavailable (auto mode, scheduled run, headless agent) or the user said "automatic" / "no questions", use these defaults. Always echo the defaulted decisions in the final report so the user can override.

Decision point	Default	Rationale
Which ideas to save (Brainstorm Phase 1)	Save 2 ideas: the most diverse pair (1 UX/polish + 1 new capability), preferring Low/Medium complexity.	Diversity beats volume; low complexity proves the workflow without committing to large work.
Duplicate detected (Brainstorm Phase 2)	Skip writing the new file.	Never overwrite user-authored notes silently.
Cleanup deletion (Cleanup Phase 2)	Refuse — report candidates only, do not delete.	Deletion is unrecoverable; require an explicit follow-up command.