product-brainstorming-with-team

Product Brainstorming With Team

Overview

Multi-agent collaborative forum for feature research. Dispatches parallel expert agents, maintains shared meeting notes, runs complexity-based discussion loops (LOW=3, MEDIUM=5, HIGH=7) with human input on all decision points ( tensions, architecture, scope, priorities).

This is NOT single-agent brainstorming. You MUST spawn actual subagents for each perspective.

Who runs this skill: Only the main thread (meeting owner) can execute this workflow. The main thread has access to the Task tool for spawning subagents. If you are a subagent (dispatched by another agent), you do NOT have this capability - report back to the main thread instead.

The Process

digraph forum_process {
    rankdir=TB
    node [shape=box, style=rounded]

    clarify [label="1. EXHAUSTIVE CLARIFICATION\n(see details below)\nDO NOT PROCEED until\nall context is gathered"]
    classify [label="2. CLASSIFY COMPLEXITY\nLOW/MEDIUM/HIGH\nDetermines loop count"]
    create_notes [label="3. CREATE MEETING NOTES\ndocs/epics/<feature>/meeting-notes.md\nShared memory for all agents"]
    dispatch [label="4. PARALLEL DISPATCH\nSpawn ALL expert agents\nin single message"]
    collect [label="5. COLLECT & SYNTHESIZE\nRead agent outputs\nUpdate meeting notes"]
    decisions [label="6. IDENTIFY DECISION POINTS\nTensions, architecture choices,\nscope tradeoffs, priorities"]
    human [label="7. ASK HUMAN\nPresent ALL decision points\nGet direction on each"]
    loop_check [shape=diamond, label="Reached loop\ncount for\ncomplexity?"]
    redispatch [label="8. REDISPATCH ALL\nShare human input +\ndecisions with agents"]
    write_prd [label="9. WRITE PRD\ndocs/epics/<feature>/prd.md"]

    clarify -> classify -> create_notes -> dispatch -> collect -> decisions -> human -> loop_check
    loop_check -> redispatch [label="no"]
    redispatch -> collect
    loop_check -> write_prd [label="yes"]
}

Step 1: Exhaustive Clarification (CRITICAL)

DO NOT MAKE ASSUMPTIONS. Ask until you're 100% clear.

Before dispatching ANY agents, you MUST gather complete context from the human. Incomplete understanding leads to wasted agent cycles and wrong recommendations.

Required Context Categories

Category	Questions to Ask
What exists today?	Current system? Tech stack? User base size? Known pain points?
Prior decisions	What's already been decided? What approaches were rejected and why?
Scope boundaries	What's explicitly IN scope? What's OUT of scope?
Success criteria	How will you know this succeeded? Metrics?
Constraints	Budget? Timeline? Team skills? Compliance requirements?
Dependencies	What other systems/teams are involved? Blockers?

Clarification Rules

1. Ask in rounds - Don't ask 20 questions at once. Ask 3-5, get answers, ask follow-ups.
2. Confirm understanding - Summarize what you heard. "So to confirm: [X, Y, Z]. Correct?"
3. Probe vague answers - "Use existing backend" → Which services? What database? What auth exists today?
4. No implicit assumptions - If you're assuming something, ASK instead.

Red Flags - You're Making Assumptions If:

• You're guessing which technologies/providers to consider
• You're imagining what the current system looks like
• You're assuming users have/don't have certain capabilities
• You're filling in blanks the human didn't specify
• You think "they probably mean X" without confirming

STOP. Ask the human. Then proceed.

When to Move to Step 2

Only proceed to classify complexity when you can answer ALL of these:

• I know exactly what problem we're solving
• I know what exists today (not assumed)
• I know what's been tried/rejected before
• I know the explicit scope boundaries
• I have confirmed my understanding with the human

---

Step 2: Classify Complexity (DETERMINES LOOP COUNT)

After clarification, classify the feature's complexity. This determines how many discussion loops you MUST run.

Complexity Levels

Level	Loop Count	Criteria
LOW	3 loops	Single domain, clear requirements, minimal integrations, well-understood tech
MEDIUM	5 loops	Cross-domain, some ambiguity, multiple integrations, new tech evaluation
HIGH	7 loops	Multi-system, novel domain, regulatory/compliance, significant architecture changes

Classification Checklist

Mark as HIGH if ANY of these apply:

• Involves regulatory/compliance requirements (GDPR, HIPAA, SOC2, etc.)
• Requires significant architecture changes
• Touches 3+ existing systems
• Involves AI/ML with novel use cases
• Has security-critical components (auth, payments, PII)
• Unclear competitive landscape requiring deep research

Mark as MEDIUM if ANY of these apply (and not HIGH):

• Crosses 2+ team boundaries
• Requires evaluating new technologies
• Has dependencies on external systems/APIs
• Involves user-facing changes affecting >10% of users
• Needs market research for positioning

Mark as LOW only if:

• Single domain, single team
• Well-understood requirements
• Standard tech stack
• Internal tooling or low-risk features

When in doubt, classify UP. Better to have more discussion than miss critical issues.

Document Classification

In meeting notes, record:

## Complexity Classification: [HIGH/MEDIUM/LOW]

**Reason:** [1-2 sentence justification]
**Required Loops:** [3/5/7]

---

Expert Agents (MUST Dispatch All)

Agent	Model	Focus	Required Tools
Product Manager	opus	Phases, MVP scope, roadmap fit	-
Product Marketer	opus	Positioning, market fit, competition analysis	MUST use WebSearch/WebFetch
Principal Engineer	opus	Architecture, tech stack, scale	-
Security Specialist	opus	Vulnerability Assessment, Compliance, data protection, risks	-
UI/UX Specialist	opus	Usability, adoption, accessibility	-
Platform Engineer	opus	Deployment, maintenance, operations	-
Researcher	opus	Technology developments, Case studies, Current best practices and standards	MUST use WebSearch/WebFetch
ML/AI Engineer	opus	Only if AI/ML is involved	-

Web Research Requirements (CRITICAL)

Product Marketer and Researcher MUST use WebSearch and WebFetch tools. Markets change daily. Static knowledge is insufficient.

Product Marketer must research:

• Current competitor offerings and pricing
• Recent market trends and shifts
• Customer sentiment and reviews of competing products
• Industry analyst reports and forecasts

Researcher must research:

• Latest technology developments in the domain
• Recent case studies and implementations
• Current best practices and standards
• Emerging solutions and startups in the space

In agent prompts, EXPLICITLY instruct:

You MUST use WebSearch to research current market conditions.
Do NOT rely solely on training data - the landscape changes daily.
Include URLs and dates for all market claims.

Red Flag: If Product Marketer or Researcher output contains NO web citations, their analysis is stale. Redispatch with explicit web research instructions.

IRON Rules

You MUST follow these rules. No exceptions.

1. EXHAUST clarification BEFORE dispatch - No assumptions. Ask until 100% clear.
2. CLASSIFY complexity and commit to loop count - LOW=3, MEDIUM=5, HIGH=7. No shortcuts.
3. ACTUALLY dispatch subagents - Use Task tool with parallel calls in ONE message
4. Create shared meeting notes file FIRST - Agents read this for context
5. Run the FULL loop count for complexity - Not 1-2 fewer "because we covered it."
6. Ask human on ALL decision points - Not just tensions (MUST use AskUserQuestion tool)
7. Product Marketer & Researcher MUST use WebSearch/WebFetch - No stale market data.
8. Update meeting notes after each round - Shared memory must stay current

Human Decision Points (CRITICAL - Not Just Tensions)

Ask the human about ALL of these, not just tensions:

Decision Type	When to Ask	Example
Tensions	Experts disagree on approach	"Engineer says build custom, Platform says use SaaS. Which direction?"
Architecture Choices	Multiple valid technical paths	"Microservices vs monolith? Event-driven vs REST?"
Scope Tradeoffs	Feature vs timeline conflicts	"Full feature takes 3 months, MVP takes 1. Which scope?"
Priority Conflicts	Multiple features competing	"Security hardening vs new features first?"
Resource Allocation	Team/budget decisions	"Hire specialist or train existing team?"
Build vs Buy	Make or purchase decision	"Build auth system or use Auth0/Okta?"
Risk Tolerance	Uncertainty in approach	"New tech is faster but unproven. Accept risk?"
User Segmentation	Who to serve first	"Enterprise or SMB first? Different UX needs."

Decision Point Collection Process

After each round, categorize ALL open questions:

## Round N Decision Points

### Tensions (Expert Disagreements)

- [List conflicts between expert perspectives]

### Architecture Decisions Needed

- [List technical choices requiring human input]

### Scope/Priority Tradeoffs

- [List feature vs timeline/resource conflicts]

### Risk Decisions

- [List uncertain approaches needing risk tolerance input]

Use AskUserQuestion for EACH category with options. Don't batch all decisions into one question - group by type for clarity.

Red Flags - STOP and Fix

• Making assumptions about what exists, what's been tried, or what's in scope
• Planning to "simulate" or "represent" perspectives yourself
• Skipping subagent dispatch because "it's more efficient"
• Resolving tensions without asking the human
• Only asking about tensions - architecture choices, scope tradeoffs, etc. also need human input
• Stopping before reaching the loop count for your complexity level
• Product Marketer or Researcher output has no web citations - redispatch with explicit web instructions
• Forgetting to create/update meeting notes file
• Proceeding with vague answers like "use existing backend" without probing
• Classifying as LOW to reduce loops - when in doubt, classify UP

All of these mean: You're violating the workflow. Fix it.

Parallel Dispatch Template

In a SINGLE message, dispatch ALL agents:

[Task tool: Product Manager agent]
[Task tool: Product Marketer agent]
[Task tool: Principal Engineer agent]
[Task tool: Security Specialist agent]
[Task tool: UI/UX Specialist agent]
[Task tool: Platform Engineer agent]
[Task tool: Researcher agent]

Each agent prompt MUST include:

1. The feature being explored
2. Path to meeting notes file to read
3. Their specific focus area
4. Request for 3-5 bullet points of concerns/ideas

Meeting Notes Template

Create at docs/epics/<feature>/meeting-notes.md:

# <Feature> - Meeting Notes

## Goal

[From clarification step]

## Audience

[From clarification step]

## Constraints

[From clarification step]

## Complexity Classification: [HIGH/MEDIUM/LOW]

**Reason:** [1-2 sentence justification]
**Required Loops:** [3/5/7]

---

## Round 1: Initial Perspectives

### Product Manager

[Agent output]

### Product Marketer

[Agent output]

### Principal Engineer

[Agent output]

### Security Specialist

[Agent output]

### UI/UX Specialist

[Agent output]

### Platform Engineer

[Agent output]

### Researcher

[Agent output]

## Round 1 Decision Points

### Tensions (Expert Disagreements)

- [Tension 1]
- [Tension 2]

### Architecture Decisions Needed

- [Decision 1]

### Scope/Priority Tradeoffs

- [Tradeoff 1]

### Risk Decisions

- [Risk 1]

## Human Input (Round 1)

[Record human's decisions for EACH category above]

---

## Round 2: Refined Perspectives

[Continue pattern for each round...]

Output

After completing all loops for your complexity level, write PRD to docs/epics/<feature>/prd.md:

# <Feature> - Product Requirements Document

## Executive Summary

[1-2 paragraphs]

## Goals & Success Metrics

[From discussion]

## User Stories

### Story 1: [Short Title]

**As a** [user type/persona],
**I want** [capability/feature],
**So that** [benefit/value].

**Acceptance Criteria:**

- [ ] Given [precondition], when [action], then [expected result]
- [ ] Given [precondition], when [action], then [expected result]
- [ ] [Edge case handling]
- [ ] [Error state handling]

**Priority:** [P0/P1/P2]
**Complexity:** [S/M/L/XL]

---

### Story 2: [Short Title]

**As a** [user type/persona],
**I want** [capability/feature],
**So that** [benefit/value].

**Acceptance Criteria:**

- [ ] Given [precondition], when [action], then [expected result]
- [ ] ...

**Priority:** [P0/P1/P2]
**Complexity:** [S/M/L/XL]

---

[Continue for all stories...]

## Technical Approach

[From Principal Engineer + Platform Engineer]

## Security & Compliance

[From Security Specialist]

## UX Requirements

[From UI/UX Specialist]

## Go-to-Market Considerations

[From Product Marketer - MUST include web research citations]

## Market Research Summary

[From Researcher - MUST include web research citations with URLs and dates]

## Phases

[From Product Manager]

## Open Questions

[Unresolved from discussion]

## Appendix: Meeting Notes

See meeting-notes.md for full discussion history.

User Story Requirements (CRITICAL)

Every user story MUST have:

1. User persona - Who benefits? (admin, end user, developer, etc.)
2. Capability - What can they do?
3. Value - Why does it matter?
4. Acceptance Criteria - How do we know it's done?

• Use Given/When/Then format
• Include happy path AND edge cases
• Include error states

5. Priority - P0 (must have), P1 (should have), P2 (nice to have)
6. Complexity - S (hours), M (days), L (week), XL (weeks)

Red Flag: If a PRD has vague user stories like "Users can login" without acceptance criteria, it's incomplete. Add specific criteria before finalizing.

Common Rationalizations (Don't Fall For These)

Excuse	Reality
"I can represent all perspectives myself"	No. Different agents bring different reasoning. Dispatch them.
"Simulating is more efficient"	Efficiency isn't the goal. Diverse perspectives are.
"One loop covered everything"	It didn't. Run your complexity's loop count. Insights emerge over iterations.
"This is simple, LOW complexity"	Did you check the criteria? When in doubt, classify UP.
"The tensions are obvious, no need to ask"	Ask anyway. Human decides, not you.
"Only tensions need human input"	Wrong. Architecture choices, scope tradeoffs, priorities ALL need human input.
"My training data has market info"	Markets change daily. Product Marketer and Researcher MUST use WebSearch.
"Web research is optional"	No. If no web citations, the analysis is stale. Redispatch.
"Meeting notes file is overhead"	It's required shared memory. Create it.
"I don't have the Task tool"	You might be a subagent. Only main thread runs this skill. Report back.
"I have enough context to start"	Do you? Review the checklist. Can you answer ALL items without guessing?
"The user gave me the basics"	Basics aren't enough. Probe deeper. What exists today? What's been tried?
"I'll clarify as tensions emerge"	Too late. Agents will waste cycles on wrong assumptions. Clarify NOW.
"Asking too many questions is annoying"	Wasting time on wrong direction is worse. Ask until clear.
"User stories don't need acceptance criteria"	Yes they do. Without criteria, how do you know it's done? Add them.