Phx Brainstorm

Brainstorm — Adaptive Requirements Gathering

Interactive interview → research → synthesis loop. Produces structured

interview.md that $elixir-phoenix:phx-plan detects and consumes (skipping clarification).

Usage

$elixir-phoenix:phx-brainstorm Add some kind of notification system
$elixir-phoenix:phx-brainstorm Improve authentication security
$elixir-phoenix:phx-brainstorm                    # starts with open question

Workflow

$elixir-phoenix:phx-brainstorm {topic}
    |
    v
[INTERVIEW] ←──────────────────┐
    |                           |
    v (sufficient OR user exit) |
[DECISION POINT]                |
    ├─ Research ──→ [RESEARCH] ─┘
    ├─ Continue interview ──────┘
    ├─ Make a plan ──→ STOP (suggest $elixir-phoenix:phx-plan {slug})
    ├─ Store & exit ──→ STOP (artifacts saved)
    └─ Discuss ──→ freeform ──→ [DECISION POINT]

Phase 1: Adaptive Interview

Create .claude/plans/{slug}/ directory. Start asking ONE question at a time.

Coverage Dimensions

Track coverage across 6 dimensions (0=uncovered, 1=partial, 2=sufficient).

Ask Scope early — for "optimize X" topics, ask about boundaries (upstream

OK? Local-only? CI vs dev?) before research, not during.

| Dim | Target | Sufficient signal |

|-----|--------|-------------------|

| What | Specific behavior/features | Concrete verbs, not "some kind of" |

| Why | Problem solved, user need | Clear benefit stated |

| Scope | In/out boundaries | Explicit exclusions stated |

| Where | Modules, contexts, pages | File paths or context names mentioned |

| How | Approach, constraints | At least one concrete constraint |

| Edge | Error states, scale, auth | 2+ edge cases identified |

Interview is "sufficient" when total score >= 8 out of 12.

Context-Aware Questioning

Before each question, run a brief codebase scan on topics the user mentioned:

• User mentions a topic (e.g., "notifications") → run Grep/Glob for related patterns
• Use scan results to ground your next question in what actually exists
• Unknown/niche topic → suggest research pause before continuing

Signal Detection

• Vague answer ("maybe", "not sure") → probe deeper on same dimension
• Niche topic mentioned → "This involves {X}. Want me to research it first?"
• Detailed answer covering 3+ dimensions → mark all covered, advance
• No new coverage for 2 consecutive questions → suggest moving to Decision Point

Phase 2: Decision Point

MANDATORY: Write interview.md FIRST, then use AskUserQuestion.

Never let the conversation flow past this point without a formal choice.

• Write current state to .claude/plans/{slug}/interview.md
• Show coverage summary: "Coverage: What 2/2 | Why 2/2 | Scope 1/2 | ..."
• Use AskUserQuestion with EXACTLY these options (4 max — the tool's

hard limit; the auto-added "Other" covers freeform discussion):

• Research — search codebase + internet for approaches (2 agents)
• Continue interview — ask more questions
• Make a plan — I'll suggest: $elixir-phoenix:phx-plan .claude/plans/{slug}/interview.md
• Store & exit — save everything, come back later

• Wait for user response. Do NOT proceed without explicit choice

AskUserQuestion discipline: decisions only, never narration or

rhetorical check-ins. Every option states concrete impact (what happens,

what it costs) so the user can pick without follow-up questions.

Phase 3: Research (Diverge → Evaluate → Converge)

First cycle: MAX 2 agents — keep it fast (~2-3 min). Spawn in ONE

Tool Use block with run_in_background: true:

• phoenix-patterns-analyst: "How does this codebase handle {topics}?"

Write to .claude/plans/{slug}/research/codebase-scan.md

• web-researcher: "Elixir/Phoenix approaches to {topics}"

Return 500-word summary

Do NOT spawn additional specialist agents in the first cycle.

If user wants deeper investigation, they pick "More research" at the

next Decision Point — then spawn focused agents for specific questions.

Evaluate — for each approach found:

• Thesis: why it works for THIS codebase
• Antithesis: why it might NOT work (scale, complexity, pattern conflicts)

Converge — present 2-3 approaches with honest trade-offs.

Do NOT recommend one. Return to Decision Point (AskUserQuestion).

See references/research-integration.md for details.

Iron Laws

• NEVER auto-transition to $elixir-phoenix:phx-plan — always present as option, let user choose
• ONE question at a time — never dump a question list
• Always write artifacts — interview.md is the contract with $elixir-phoenix:phx-plan
• Scan codebase between questions — every question must be context-aware
• AskUserQuestion at EVERY decision point — never flow past without formal choice.

This is the most critical law. After interview, after research, after discuss — ALWAYS

present options via AskUserQuestion. Never let conversation skip the checkpoint

• STOP after presenting options — do not proceed without user input
• MAX 2 agents in first research cycle — deeper dives are subsequent cycles.

User picks "More research" to go deeper, not the skill

Integration

$elixir-phoenix:phx-brainstorm ──→ interview.md ──→ $elixir-phoenix:phx-plan (skips clarification)
                                 ──→ $elixir-phoenix:phx-plan --existing (deepens)
                                 ──→ stored for later session

Position: optional upstream of $elixir-phoenix:phx-plan in workflow cycle.

References

• references/interview-techniques.md — coverage scoring,

question templates, scan patterns, signal detection, interview.md format

• references/research-integration.md — diverge-evaluate-converge,

agent spawn templates, approach presentation format