AI initializes the session with a 3-question interactive setup then presents a technique selection menu.

• 1
  AI asks 3 setup questions to define session scope

  • topic — main topic of the brainstorm
  • goal — broad exploration or focused ideation (deep)
  • capture session — whether to save the output document (default Yes)
• 2
  AI offers 4 technique selection modes

  • Mode 1 — User-selected: user browses library of ~60 techniques and picks
  • Mode 2 — AI-recommended: AI analyzes context, suggests 3-4 best fits with one-line rationale
  • Mode 3 — Random: AI picks randomly (good for breaking patterns)
  • Mode 4 — Progressive flow: moves sequentially from broad → narrow through multiple techniques
• 3
  User confirms choice before moving to step 02 — no auto-progress

• ·
  No suitable technique → session becomes unfocused or falls into anchor bias
• ·
  Each technique opens a different angle of attack on the problem space
• ·
  SCAMPER is effective when improving existing solutions (Substitute / Combine / Adapt / Modify / Put to other use / Eliminate / Reverse)
• ·
  Reverse Brainstorming is effective when stuck — ask "how to create the problem?" instead of "how to solve it?"
• ·
  Six Thinking Hats is effective when balancing multiple perspectives is needed (facts / emotions / risks / benefits / creative / process)
• ·
  Mind Mapping is effective for broad exploration without a specific scope

AI loads the technique library (~60 entries) from the internal knowledge base then routes by mode.

• 1
  User-selected mode: AI displays the full list with category labels (divergent / convergent / lateral / structured)
• 2
  AI-recommended mode: AI analyzes 2 axes

  • user context (industry, problem type, stage)
  • strengths of each technique (internal matching matrix)
  • returns top 3-4 techniques with a one-line justification
• 3
  After user picks, AI confirms with a summary: "OK, we'll use [technique] for [goal]. Ready to start?"

AI opens the generation session, prompts according to the chosen technique, generates ideas in batches of 10-15 ideas/round, targeting 100+ ideas total.

• 1
  AI prompts the user according to the technique template (e.g., SCAMPER: "What can we Substitute in the current design?")
• 2
  Each round is timeboxed ~5-10 minutes, AI applies 3 principles

  • defer judgment — no evaluation of any idea during this round
  • build on others — combine existing ideas into new ones
  • wild ideas welcome — outrageous ideas are encouraged
• 3
  After each round, AI asks for brief feedback ("good direction?") then continues
• 4
  AI does not self-generate ideas but pulls ideas out of the user via probing questions
• 5
  Track count continuously, target 100+ ideas before moving to convergence

• ·
  Per creativity research: the first 20 ideas are usually obvious, ideas 50-100 are the "golden" zone
• ·
  Early critique kills divergent thinking — weak ideas still have value as stepping stones to good ones
• ·
  Quantity-first breaks self-censorship — users verbalize minor ideas more easily when they know filtering comes later
• ·
  Timeboxed rounds maintain energy — open-ended sessions tend to degrade in quality after 30 minutes

• ·
  AI uses probing questions such as: "What if cost was zero?", "What would Tesla do?", "What's the opposite approach?"
• ·
  AI uses analogical prompting: "What if this were like X industry?" to break domain anchoring
• ·
  Each idea is recorded with a 1-line description, no expansion required — speed > polish at this stage
• ·
  AI counts silently and signals when milestones are reached ("50 ideas reached, keep going")

After every ~10 ideas, AI actively shifts the lens to break patterns and open new idea categories.

• 1
  AI tracks session metadata: idea count, which technique, which cluster is dominating
• 2
  When threshold is reached (every 10/20/30 ideas), AI self-pivots using one of 3 methods

  • Persona shift: "Now think as a CFO / kid / hacker / regulator"
  • Hat shift: switch Six Thinking Hat (white = facts → red = emotions → black = risks → green = creative...)
  • Question inversion: "How would we make this feature FAIL?"
• 3
  AI does not ask permission but declares "pivot now — let's look from X angle" and continues
• 4
  After a pivot, new-round ideas are often more outlier → flag separately to avoid early merging

• ·
  Anchoring bias: the brain anchors to the initial framing — pivots break that anchor
• ·
  Availability heuristic: the next idea tends to be close to the last one → homogeneous clusters
• ·
  Functional fixedness: only seeing 1 use case per thing — persona shifts open alternative use cases
• ·
  Active pivots (without waiting for user request) create structured serendipity

• ·
  AI uses an internal perspective rotation matrix: 6 personas × 3 timeframes × 4 constraint levels
• ·
  Persona shifts in a 60-minute session: 3-4 different personas, ~15-20 ideas per persona
• ·
  Standard Six Hats sequence: white (data) → red (feelings) → black (risk) → yellow (benefit) → green (creative) → blue (process)
• ·
  Question inversion technique from Reverse Brainstorming: list ways to cause the problem, then negate them

Convergence mode begins. AI groups ideas into clusters by affinity, names categories, and deduplicates.

• 1
  AI scans all 100+ ideas and proposes 5-9 initial clusters
• 2
  Cluster names use "categories everyone understands"

  • e.g., UX improvements, Revenue models, Tech debt, Compliance
  • avoid jargon or team-internal names
• 3
  User approves or renames each cluster
• 4
  AI assigns each idea to exactly 1 cluster
• 5
  Outlier ideas (one-of-a-kind, don't fit any cluster)

  • placed in a separate parking lot
  • never forced into a cluster (preserves originality)
• 6
  Duplicate ideas merged by content (merge count → signals cluster importance)

• ·
  100+ spread-out ideas are noise — not actionable, not communicable
• ·
  Clustering turns noise into signal — reveals themes the session naturally gravitated toward
• ·
  "Populous" themes are usually the most investment-worthy areas of the problem space (revealed preference)
• ·
  "Sparse" themes also carry information — gaps may be blind spots or less important areas
• ·
  Per Miller's Law (7±2): human short-term memory holds ~5-9 items — clusters in this range are easier to retain

• ·
  AI uses affinity mapping protocol: physical metaphor (sticky notes on a wall) even in digital form
• ·
  Clustering is conceptual — by purpose/audience/mechanism, not by similar wording
• ·
  After clustering, count ideas/cluster → bar chart visual so the user can see distribution
• ·
  Parking lot outlier ideas are often material for a future epic — do not discard

AI applies a prioritization framework, scores ideas, and selects the top 5-10 to carry into Phase 2.

• 1
  AI offers 2 prioritization frameworks

  • Impact/Effort 2×2 matrix: quick wins / big bets / fill-ins / thankless tasks
  • MoSCoW: Must have / Should have / Could have / Won't have
• 2
  User selects framework (default: Impact/Effort — more visual)
• 3
  Scoring process

  • Impact (1-5): "if shipped, how much user/business value does this deliver?"
  • Effort (team-weeks): "how long to build and ship?"
  • AI proposes initial scores, user adjusts
• 4
  Plot ideas on the visual matrix
• 5
  Output: top 5-10 ideas with short rationale (1-2 lines each)
• 6
  Quick wins (high impact + low effort) flagged separately for immediate action

• ·
  A great brainstorm without prioritization = a dead document
• ·
  Prioritization forces explicit trade-offs and prevents cherry-picking later
• ·
  Scoring surfaces hidden disagreement: if 2 people score very differently, different assumptions need to be clarified
• ·
  Quick wins shipped early create momentum + early evidence for the Phase 2 PRD
• ·
  Big bets need additional research/validation before commitment — flag to avoid hasty decisions

• ·
  AI uses Impact/Effort 2×2 in Eisenhower box style — 4 quadrants with pre-defined names
• ·
  Effort dimension: AI estimates using reference class forecasting — compares against similar features with known history
• ·
  Impact dimension: AI tries to quantify ("saves 2 hours/user/week") rather than using generic ratings ("high")
• ·
  Avoid vanity metrics: prioritize retention/engagement over page views/downloads

AI generates brainstorming-report.md containing the full session and saves it to _bmad-output/planning-artifacts/.

• 1
  Report sections in order

  • Setup: topic, goal, technique(s) used, date, attendees
  • Full idea list: numbered, grouped by round
  • Clusters: with count + name + member ideas
  • Prioritization matrix: visual 2×2 or MoSCoW table
  • Top picks: 5-10 ideas with rationale
  • Next steps: recommended workflow to continue
• 2
  Frontmatter attaches metadata: stepsCompleted, workflowType: brainstorming, date
• 3
  AI invokes bmad-help with arg = "brainstorming"
• 4
  bmad-help suggests next workflow based on project state

  • typically: research (validate ideas) or product-brief (document vision)
  • if ideas are already solid: go straight to create-prd

• ·
  Output artifact = single source of truth for subsequent phases
• ·
  Without a file, context is lost when changing chat sessions (core BMAD principle: "artifacts drive state")
• ·
  Teammates cannot re-derive results from a verbal summary alone
• ·
  Frontmatter metadata enables workflow continuation — re-invoking a workflow knows which steps are already done

• ·
  Append-only writing: AI writes section by section, does not regenerate the file
• ·
  File path resolves via config: {planning_artifacts}/brainstorming-{slug}-{date}.md
• ·
  bmad-help auto-invokes at the final step of every workflow — uniform design pattern
• ·
  Roadmapping syntax: next phase, next workflow, prerequisite, estimated time

AI turns a vague question into 5-8 sub-questions answerable with evidence.

• 1
  AI asks user to select one or more research scopes

  • Market — TAM/SAM/SOM, competitors, trends, pricing
  • Domain — vocabulary, expert models, regulations
  • Technical — feasibility, stack options, implementation paths
• 2
  AI applies 5 Whys to drill from the original question down to the root question

  • From each answer, AI asks 'why?' again
  • Stop at level 5 or when a researchable question is reached
• 3
  AI converts root question into 5-8 specific, evidence-answerable sub-questions
• 4
  User validates the sub-question list before AI begins gathering

• ·
  A vague question leads to wandering research — specificity creates clear coverage checklists
• ·
  Small sub-questions enable parallel investigation and a clear signal for 'done'
• ·
  Surfaces scope creep early — one original question can imply 10 sub-areas

• ·
  Use the problem framing canvas: problem / users affected / current solutions / why insufficient
• ·
  Each sub-question must be answerable with evidence, not an opinion question
• ·
  Standard format: "Are there competitors using on-device LLM for document Q&A, and what is their pricing?"

AI gathers evidence from 3+ independent source types, attaching a full citation to each finding.

• 1
  AI searches in parallel across 4 source types

  • Web search — news, blogs, industry reports
  • Official docs — specs, standards (ISO/RFCs), regulatory texts
  • Expert content — podcasts, talks, academic papers
  • Interview notes — if the user has existing user research
• 2
  Each finding is assigned a quality grade

  • Primary — regulator docs, SEC filings, official statements
  • Secondary — Reuters/Bloomberg, peer-reviewed research
  • Tertiary — Wikipedia, aggregators, blog opinions
• 3
  Each citation includes: URL + publish date + author name + verbatim quote or clear paraphrase

• ·
  Triangulation — a fact appearing in 3 independent sources = evidence; 1 source = assumption
• ·
  Mandatory citations prevent AI from hallucinating stats — must attach a specific URL
• ·
  Quality grading helps downstream synthesis weight evidence appropriately

• ·
  Desk research protocol: primary first, then secondary, tertiary only for cross-checking
• ·
  Multi-query strategy: rephrase each sub-question 2-3 different ways to cover the keyword space
• ·
  Recent-first filter: prioritize sources from the past 12 months for fast-changing topics

AI cross-references data and looks for 3 types of signal: patterns, contradictions, gaps.

• 1
  AI builds a finding matrix

  • Rows — sub-questions from step 01
  • Columns — sources gathered
  • Cells — quote/data from that source for that question; empty cell = gap
• 2
  AI scans the matrix for 3 signal types

  • Patterns — ≥3 sources agree on X → high-confidence finding
  • Contradictions — source A says X, source B says ~X → highlight and investigate
  • Gaps — empty cells in the matrix → unanswered questions
• 3
  AI does not skip contradictions or gaps — both are predictive of risk

• ·
  Pattern recognition is easy to overweight — contradictions and gaps are often overlooked but carry high value
• ·
  Contradictions are typically contested ground worth digging into
• ·
  Gaps are often unknown unknowns — predictive of risks in Phases 3-4

• ·
  Matrix rendered as a table (sub-questions × sources); empty cells naturally surface as gap indicators
• ·
  Pattern threshold: ≥3 independent sources = pattern; 2 = weak signal; 1 = assumption
• ·
  Contradictions: AI presents both sides with evidence weight, does not pick a winner

AI synthesizes into 3-5 actionable insights, quantifying when possible.

• 1
  Each insight has the format

  • Claim — 1 line, specific, quantified when data is available
  • Evidence pointer — cite which source numbers in the matrix
  • Confidence level — high / medium / low
• 2
  AI lists 3-5 risk flags cross-category

  • Market — timing, demand, competition
  • Technical — feasibility, scalability, vendor
  • Regulatory — compliance, data, jurisdiction
• 3
  Output: 1 document readable in 5 minutes by a PM/founder

• ·
  Synthesis ≠ summary — synthesis creates new understanding by combining ('because X and Y, therefore Z')
• ·
  Quantifying forces precision — 'growing fast' could be 5% or 50%, a critical difference
• ·
  Confidence level helps PM weight insights when they conflict with gut feeling

• ·
  Pyramid Principle (Minto): conclusion-first, supporting evidence below
• ·
  Confidence calibration: high = 3+ primary sources; medium = 2 sources; low = 1 + inference
• ·
  Each insight tagged with evidence count (X sources) for quick reviewer judgment

AI provides evidence-based recommendations — each with action, rationale, owner, metric.

• 1
  Each recommendation has the format

  • Given — evidence from insights
  • Recommend — specific action
  • Because — reasoning linking evidence → action
  • Owner — PM / Tech Lead / Founder
  • Success metric — how to know it worked
  • Risk if skipped — consequence of inaction
• 2
  AI avoids opinion-only — each recommendation must trace back to evidence from steps 03/04
• 3
  If evidence is weak, AI flags low-confidence recommendation instead of pretending it's solid

• ·
  Research without recommendations = trivia — PMs need decisions, not facts
• ·
  Given/Recommend/Because format makes recommendations defensible in meetings with stakeholders
• ·
  Risk if skipped creates a 'cost of inaction' — helps prioritize against other proposals

• ·
  Argumentation theory: claim + grounds (evidence) + warrant (reasoning)
• ·
  Ranking: impact × confidence × urgency
• ·
  AI offers 2-3 alternative paths when there is genuine uncertainty instead of forcing 1 answer

AI generates research-findings.md with full citations in APA/Chicago format.

• 1
  Report structure

  • Executive summary — 1 paragraph
  • Methodology — scope, sources, date range
  • Sub-questions + evidence — 1 section/question
  • Insights — 3-5 with confidence levels
  • Recommendations — ranked with rationale
  • Full bibliography — alphabetical, complete citations
• 2
  Frontmatter: dates, methodology, scope, workflowType: research
• 3
  AI invokes bmad-help → suggest next workflow based on strength of findings

• ·
  Citations ensure reproducibility — PM can trace every claim back to its source
• ·
  Methodology section enables scope critique — reviewer can see if key sources are missing
• ·
  Artifacts drive state — without a file, context is lost when switching chats

• ·
  Format: APA (Author, Year, Title, URL) or Chicago (with footnotes)
• ·
  Append-only writing — AI writes each section, does not regenerate the file
• ·
  End of file includes a 'Last updated' timestamp to track freshness

AI loads prior artifacts and scopes the brief for the target audience.

• 1
  AI checks the workspace for Phase 1 artifacts

  • brainstorming-report.md
  • research-findings.md
  • prfaq-{project}.md
• 2
  If found: AI extracts relevant content — no re-derivation
• 3
  AI asks 3 scoping questions

  • Audience — founder / engineering / mixed / stakeholder external
  • Decision driven — what decision does this brief inform?
  • Level of detail — 1-2 pages (preferred) or deeper
• 4
  AI propose outline → user adjust → commit

• ·
  Brief without scope → becomes a PRD lite — long, rambling
• ·
  A proper brief must be 1-2 pages, stakeholders read it in 5 minutes for a go/no-go decision
• ·
  Extracting (not re-deriving) saves time and maintains consistency with earlier artifacts

• ·
  Default audience: mixed (founder + early team) if user does not specify
• ·
  Artifact extraction: AI scans frontmatter + headings, lists relevant sections for user confirmation
• ·
  Outline proposed as bullet list for user to adjust before writing

AI co-drafts Vision (5-10 years) and Mission (1 sentence, what-for-whom-how).

• 1
  Vision drafting

  • AI ask: 'What does world look like in 5-10 years if you succeed?'
  • AI propose 2-3 draft options
  • User selects + iterates until no more words are needed
• 2
  Mission drafting with the Geoffrey Moore template

  • 'For [user], who [pain], our [product] is [category] that [value]. Unlike [alternative], we [differentiator].'
  • Compress down to the final 1-2 sentences
• 3
  Validate: vision is distant from tactics; mission is close to tactics

• ·
  Vision alone = a dream with no daily direction
• ·
  Mission alone = labor without purpose, losing long-term direction
• ·
  The two are complementary and serve as an anchor whenever feature creep occurs

• ·
  Opposite test: if the opposite is not plausible → vision is too bland, redo
• ·
  Read aloud test: if it sounds generic / applies to any startup → re-draft
• ·
  Vision/Mission must be finalized before step 03 — all future decisions reference back to them

AI sketches 1-3 personas: demographics, JTBD, pain, current alternative, success criteria.

• 1
  Each persona format

  • Demographics — age, role, location, tech savviness
  • JTBD — 'When [situation], I want to [motivation], so I can [outcome]'
  • Primary pain — 1-2 specific lines
  • Current alternative — how they solve it today (including 'do nothing')
  • Success criteria — measurable outcome if pain resolved
• 2
  AI does not merge all users into 1 'average user'
• 3
  Validate with user

  • 'Do you know a real person who matches this persona?'
  • If not → persona may be fiction, redo

• ·
  Average user is fiction — real users have different jobs/pain
• ·
  Products trying to serve everyone usually serve no one
• ·
  JTBD framework focus on motivation rather than demographic stereotype

• ·
  AI uses Clayton Christensen's JTBD framework
• ·
  AI uses the empathy map as a supplementary tool: says / thinks / does / feels
• ·
  Also define the anti-persona — who we explicitly do not serve

AI writes a Problem Statement specific enough to be falsifiable.

• 1
  4 components of the problem statement

  • What — specific problem, not a solution disguised as a problem
  • Who — who is affected, link to personas
  • Why important — frequency × severity × visibility
  • Why now — timing: regulation, tech enablement, market shift
• 2
  AI avoids solution-in-the-problem (e.g., 'we need a chatbot' = solution, not a problem)
• 3
  Statement must be falsifiable: can be refuted by evidence

• ·
  Vague problem → vague solution → shipping features that don't address the core issue
• ·
  The problem statement is a contract between the team and reality
• ·
  Being specific enables measuring real success, not just 'feature shipped'

• ·
  PAIN framework: frequency × severity × visibility
• ·
  Falsifiability test: 'What evidence would prove this problem doesn't exist?'
• ·
  AI uses 5 Whys to dig past surface complaints to the root problem

AI defines the Value Proposition: unique value vs. alternatives, mapping pains → gains.

• 1
  AI uses the Value Proposition Canvas

  • Customer pains ← linked to problem statement
  • Customer gains — positive outcomes desired
  • Pain relievers — how product addresses each pain
  • Gain creators — how product creates each gain
• 2
  AI maps value vs. 3 alternatives: top competitors + current workaround + 'do nothing'
• 3
  Articulate differentiator: 1-2 'only we have because [moat]'

• ·
  Most products fail not because they are bad, but because they are not different enough for users to switch
• ·
  'Do nothing' is always the strongest competitor — switching cost > pain = no adoption
• ·
  Articulating the value prop early reveals: if you cannot articulate it, it may not exist

• ·
  Value Proposition Canvas by Strategyzer/Osterwalder
• ·
  Differentiation matrix: rows = features, cols = us + top 3 competitors
• ·
  AI flags features that all competitors have = table stakes, not differentiators

AI defines Success Metrics: 1-2 north star + 3-5 supporting KPIs.

• 1
  North Star Metric — 1-2 metrics that, if improved, improve the entire business

  • Airbnb = nights booked; Spotify = time listening; Slack = messages sent
• 2
  Supporting KPIs per the HEART framework

  • Happiness — NPS, satisfaction
  • Engagement — DAU/MAU, session frequency
  • Adoption — signup conversion, time-to-first-value
  • Retention — D7/D30, cohort curves
  • Task success — completion rate, error rate
• 3
  Separate timeframes

  • Short-term (3-6 months) — leading indicators
  • Long-term (1+ year) — lagging indicators

• ·
  Metrics are a contract between the team and the board
• ·
  Vanity metrics (page views, downloads) are easy to grow but do not correlate with business value
• ·
  The North Star aligns the entire team around a single number

• ·
  OKR template + HEART framework by Google
• ·
  Each metric: baseline (current) + target (12 weeks) + measurement source
• ·
  AI flags a warning if a metric has no baseline/target → 'vague metric'
• ·
  Layer North Star + 2-3 leading indicators — avoid putting all expectations on a single number

AI lists Constraints: budget, timeline, team, technical, regulatory — categorized as MUST/SHOULD/MAY.

• 1
  AI asks the user about 5 types of constraints

  • Budget — dollar amount, runway
  • Timeline — launch date, milestones
  • Team — size, skills, hiring plan
  • Technical — must integrate with X, must use Y stack
  • Regulatory — HIPAA / GDPR / SOC2 / PCI / industry-specific
• 2
  Each constraint categorized into 3 tiers

  • MUST — deal-breaker if violated
  • SHOULD — preferred
  • MAY — nice-to-have
• 3
  AI flags conflicts between 2 MUST constraints — must be resolved before committing

• ·
  Constraints define solution space — not listing them = assuming they don't exist = surprise mid-sprint
• ·
  A constraint can become a differentiator (e.g., 'on-device' due to privacy constraint becomes a USP)
• ·
  MUST constraint conflicts must be resolved before committing — not mid-development

• ·
  Constraint analysis matrix: type × tier × source
• ·
  AI asks regulatory questions per industry: healthcare → HIPAA, finance → SOC2/PCI, EU → GDPR
• ·
  Brand constraints are often forgotten: voice, visual identity, tone restrictions

AI saves product-brief.md and clearly hands off to Phase 2.

• 1
  Final document structure

  • Executive summary — 1 paragraph (written last, placed first)
  • Vision + Mission
  • Target users — 1-3 personas
  • Problem statement
  • Value proposition + differentiation
  • Success metrics — north star + KPIs
  • Constraints — tier categorized
• 2
  Frontmatter: workflowType: product-brief, stepsCompleted complete
• 3
  AI invokes bmad-help → recommends next: create-prd or prfaq

• ·
  The brief is the input for the PRD — the PRD source-extracts the brief automatically, no re-explaining
• ·
  Executive summary written last but placed first per the Pyramid Principle
• ·
  1-2 page format enforces tight writing — quality over quantity

• ·
  Append-only writing: write each section, do not regenerate
• ·
  File path: {planning_artifacts}/product-brief.md
• ·
  bmad-help auto-invoked at the final step — uniform design pattern for all workflows

AI writes a hypothetical press release for launch day in AP/Reuters style.

• 1
  PR structure (~1 page)

  • Headline — 8-12 words, attention-grabbing
  • Subhead — 1 sentence expand on headline
  • Lead paragraph — who/what/when/where/why
  • Body — 2-3 paragraphs detail
  • Customer quotes — 2-3 concrete, specific quotes (e.g., 'saved me 4h/week')
  • Leadership quote — articulate the theory of change
  • Call-to-action — where to sign up / learn more
• 2
  AI writes as a freelance journalist — if it feels boring, the product is not exciting enough yet

• ·
  Amazon's Working Backwards — Jeff Bezos's methodology
• ·
  If you cannot write a compelling PR, the product is not ready to build
• ·
  PR forces articulating the value proposition in user-facing language, not engineer-facing
• ·
  Customer quotes force specific outcomes — vague benefits = vague product

• ·
  Inverted pyramid: most important info top, details bottom
• ·
  Boredom test: read aloud — if it feels cringe, redo
• ·
  Customer quotes format: '[Product] saved me [specific outcome]' — not 'I love it'

AI generates External FAQ (8-12 Q/A) from the perspective of a skeptical customer.

• 1
  FAQ covers the most uncomfortable topics

  • Pricing — 'Why this expensive?', 'Free tier?'
  • Comparison — 'How different from [competitor]?'
  • Risk — 'Data privacy?', 'Vendor lock-in?'
  • Migration — 'How to migrate from current solution?'
  • Support — 'What if it breaks?', 'Refund policy?'
• 2
  The hardest questions at the top of the FAQ — not buried at the bottom
• 3
  Each answer 2-4 lines: direct + concrete, no marketing fluff

• ·
  FAQs expose uncomfortable questions the team typically avoids
• ·
  Answering before launch = ready to defend the product; unable to answer = not yet solid
• ·
  Hardest questions on top — if the team is not sure = must resolve before committing

• ·
  Devil's advocate prompting: AI generate FAQ as if hostile to product
• ·
  Answer format: direct + evidence + escape hatch if applicable
• ·
  Avoid marketing fluff — concrete numbers/comparisons rather than adjectives

AI generates the Internal FAQ for the team/board: risks, costs, assumptions, failure modes.

• 1
  Internal FAQ covers

  • Top risks — market, tech, team, regulatory, financial
  • Cost breakdown — dev cost, COGS, marketing CAC
  • Timeline assumptions — '6 months — what if it takes 12?'
  • Technical trade-offs — chosen path vs alternatives
  • 'What if we're wrong about X' — pre-mortem questions
• 2
  AI uses the pre-mortem protocol

  • Assume product fails in 12 months — tell the story of WHY
  • Top 5 failure reasons → converted into a risk register

• ·
  Internal FAQ is the pre-mortem mechanism
• ·
  Per Kahneman: pre-mortem improves forecast accuracy by 30%
• ·
  Surface hidden assumptions before committing — after committing it costs 10-100x to fix

• ·
  Each risk: probability × impact + early warning signal + mitigation
• ·
  AI prompts hardest hypotheticals: 'What if Apple does this in 6 months?'
• ·
  Risk categories: Market / Technical / Team / Regulatory / Financial

AI narrates the Customer Experience end-to-end: awareness → habit → renewal.

• 1
  5 stages with target emotion

  • Awareness — how user hears about product
  • Consideration — what convinces them to try
  • Onboarding — first 5 minutes after signup
  • First use — aha moment / value realization
  • Habitual use — daily/weekly engagement loop
• 2
  AI identifies moments of truth at each transition

  • Critical decision points where user can drop off
  • Typically at: pricing page, first-use friction, error recovery

• ·
  A static features list says 'what exists'; CX narrative says 'how it feels'
• ·
  UX problems mostly occur at transitions between stages, not in the stages themselves
• ·
  Peak-end rule (Kahneman): users remember peak emotion + end emotion → design accordingly

• ·
  Service blueprint: above-the-line (user-visible) + below-the-line (backend)
• ·
  User journey mapping with emotion curve
• ·
  Deliberately create 1 peak emotion + strong end emotion

AI stress-tests the entire PR/FAQ from 5 adversarial perspectives, patches gaps, finalizes.

• 1
  AI plays adversarial reviewer from 5 angles

  • Investor — 'TAM? Unit economics? Moat?'
  • Customer — 'Why would I switch? Risk?'
  • Engineer — 'Feasible at scale?'
  • Competitor — 'How quickly can we copy?'
  • Regulator — 'What compliance is needed?'
• 2
  Each question → grade the answer

  • Strong — keep as-is
  • Weak — patch/strengthen the related section
  • No answer — flag as gap, document for later
• 3
  Finalize prfaq-{project}.md → invoke bmad-help → suggest create-prd

• ·
  Stress test = attempting to break the product on paper before breaking it in reality
• ·
  If the PR/FAQ withstands 10 adversarial questions, the concept has resilience
• ·
  Gap documentation → research backlog for Phase 2

• ·
  AI cycles through 5 personas, 2-3 hard questions per persona
• ·
  Iterate until all answers are ≥ 'strong' or gaps are explicitly documented
• ·
  Adversarial review pattern reused in Phase 4 Code Review (04·F)