Idea Extraction

Fractal Structure Protocol

This skill writes directly to the .memory/wiki/specs/ideation/ folder using a fractal pattern — every node in the tree (surface, domain, sub-domain) follows the same structure:

{node}/
├── {node}-index.md       ← what's in this node, child listing, Role Matrix
├── {node}-cx.md          ← cross-cuts connecting this node's children
├── child-01/             ← same pattern repeats (if child is complex)
└── 01.01.01-feature.md   ← leaf node = .md file, not a folder

Key rules:

1. Always folders. Every domain, sub-domain, and even single-surface projects use folders — no flat files. One universal pattern, no exceptions.
2. Indexes at every level. Template: .factory/skills/prd-templates/references/fractal-node-index-template.md
3. CX files at every level. Template: .factory/skills/prd-templates/references/fractal-cx-template.md
4. Feature files are leaf nodes. Template: .factory/skills/prd-templates/references/fractal-feature-template.md
5. Depth is reactive. The structure grows during exploration — never pre-scaffolded beyond what's discovered.
6. Soft limit at 4 levels. If creating a level 5+ node, pause and ask the user: "This is getting unusually deep — should I promote a parent to reduce nesting, or is this depth justified?"

Super-Index

The top-level ideation-index.md uses a different template: .factory/skills/prd-templates/references/ideation-index-template.md The top-level ideation-cx.md uses: .factory/skills/prd-templates/references/ideation-crosscut-template.md

Numbering Convention

Numbers are hierarchical and dot-separated within a surface:

{domain}.{sub-domain}.{sub-sub}.{feature}
  01    .    01     .   02   .    03

• Folders use number prefix: 01.03.02-ai-assistant/
• Feature files use full number: 01.03.02.03-parts-recommendation.md
• Cross-surface references prefix with surface: web/02.01.03, desktop/01.03.02
• Numbers are stable once assigned — never renumber

Structural Classification Protocol

This classification must happen BEFORE any domain files are written. It determines the folder layout for the entire ideation phase. The classification is performed in ideate-extract (Step 1.3) and recorded in ideation-index.md.

Four Project Shapes

Detection: When to Ask vs When to Detect

Interview Questions (asked early, before domain exploration)

When the input doesn't make the project shape obvious, ask these questions immediately after the opening problem statement question — before any domain exploration or file creation:

1. "Who are the distinct user types or audiences?"

   • Single audience → likely single-surface
   • Multiple distinct audiences → likely multi-product

2. "What platforms or surfaces does this need to live on?" (web, mobile, desktop, API, CLI)

   • One platform → single-surface
   • Multiple platforms, same stack → multi-surface-shared
   • Multiple platforms, different stacks or exclusive features → multi-product

3. "Is there a primary platform that the other surfaces depend on — like a central API or web platform? Or are all surfaces independent peers?" (only if multi-product)

   • Yes, one primary → hub-and-spoke
   • No, all independent → peer

Detection Signals (for document input)

When processing a document, scan for these signals before creating any domain files:

Hub-and-Spoke Implications

When hub-and-spoke is identified:

• The hub surface owns shared API/data domains. Device History, Payments, Certification — these live INSIDE the hub surface's domain tree, not in a separate shared/ folder.
• Spoke surfaces reference hub domains via CX. Desktop's CX files say "Feature X consumes web/domain/feature via API."
• Spoke surfaces own their operational domains. A domain belongs on the surface where it is primarily experienced and operated, not where the data lives. POS is a desktop domain even though it calls the hub's Stripe API. Inventory is a desktop domain even though it syncs to the hub's database.
• Proportionality check: The hub may have more domains than any single spoke, but spokes must reflect the FULL experience their users have. If a spoke user's daily workflow involves 10+ distinct capabilities, 2-3 domains is a red flag — re-examine whether concepts were incorrectly collapsed into the hub.

Peer Mode Implications

When peer mode is identified:

• Shared domains live in shared/. shared/ is treated as a peer node with the same fractal pattern.
• Each surface owns only its exclusive features.
• Surface CX files reference shared/ domains. Both surfaces point to shared, rather than one owning it.

Classification Persistence

The classification is recorded in ideation-index.md under ## Structural Classification. All downstream steps read this section to determine where to place new nodes.

Node Classification Gate

RUN THIS GATE before creating ANY new node — domain, sub-domain, or feature.

"I discovered [thing]. What is it?"
     │
     ▼
Does it belong to an EXISTING domain or sub-domain?
     │
    YES ──► Does it have 2+ distinct capabilities that interact with each other?
     │            │
     │           YES ──► It's a SUB-DOMAIN ──► create folder inside existing parent
     │            │
     │           NO ──► It's a FEATURE ──► create .md file inside existing parent
     │
    NO ──► (Surface Placement — three questions)
              │
              Q1: "Where is this PRIMARILY experienced/operated?"
              │    (Which surface's user interacts with this most?)
              │
              Q1a: "Does this concept have a DIFFERENT primary surface
              │     for initial setup vs. day-to-day operation?"
              │     (e.g., discovered on web, used daily on mobile)
              │     │
              │     YES ──► SPLIT: setup surface owns data/API domain,
              │     │        daily-operation surface gets its own operational domain
              │     │        Both are real domains — connect via CX
              │     NO  ──► proceed with single primary surface from Q1
              │
              Q2: "Which hub APIs does it consume?"
              │    (What data/services does it call from the hub?)
              │
              ├── Primarily experienced on a SPOKE surface
              │     ──► It's a SPOKE DOMAIN ──► create in that spoke surface
              │     ──► Log hub API dependencies in CX (Q2 answer)
              │
              ├── Primarily experienced on the HUB surface
              │     ──► It's a HUB DOMAIN ──► create in hub surface
              │
              ├── No primary surface (pure API/data layer, no direct user interaction)
              │     ──► It's a HUB DOMAIN (infrastructure) ──► create in hub
              │
              ├── Setup and daily-operation on DIFFERENT surfaces
              │     ──► BOTH surfaces get domains ──► connect via CX
              │     (e.g., Consumer Accounts: web owns signup/config,
              │      mobile owns daily device management/messaging)
              │
              └── Equally used across multiple surfaces (rare)
                    ──► It's a shared domain ──► create in shared/ (peer) or hub (hub-and-spoke)

**WHEN UNCERTAIN: Ask the user.** Never assume placement.

Key principle: A domain belongs where it is PRIMARILY USED, not where the data lives. POS is a desktop domain even though payments go through the hub's Stripe API. Inventory is a desktop domain even though stock data is in the hub's database. The hub API dependency is logged as a CX reference, not as a classification signal.

Sub-Domain vs Feature Test

The key question: "Does this thing have its own internal features that interact with each other?"

Anti-Patterns

Reactive Depth Protocol

Depth Grows from Exploration

The structure is NEVER pre-scaffolded. It grows as the agent explores:

Promotion: Feature → Sub-Domain

During drilling, a feature might reveal internal complexity that warrants promotion:

1. Agent discovers the feature has 2+ interacting capabilities
2. Agent asks: "This feature has enough internal complexity to be its own sub-domain. Should I promote it?"
3. If confirmed:
   • Create {number}-{slug}/ folder
   • Create index and CX files inside
   • Split content from the old feature file into child feature files
   • Update parent index
4. Numbering stays the same — the feature number becomes the sub-domain number

Exhaustion Check (Leaf-Node Model)

Exhaustion is checked at the leaves — whatever the deepest items are in each branch:

• Old model: "All domains at DEEP"
• New model: "All LEAF NODES at DEEP or EXHAUSTED"

Status propagation rules:

• All children [EXHAUSTED] → node is [EXHAUSTED]
• All children [DEEP]+ → node is [DEEP]
• Any child below [DEEP] → node stays at its current status

Role Integration

Global Persona Definitions

Personas are defined ONCE in meta/personas.md. This file is the single source of truth.

Where Persona Info Appears

Anti-Duplication Rule

• NEVER redefine a persona outside meta/personas.md
• ALWAYS reference personas by their short name
• If a new persona is discovered → add to meta/personas.md FIRST, then reference

Input-Adaptive Modes

Before starting, classify what the user has provided and select the right mode.

Extraction Mode — Rich Input

Trigger: User provides substantial existing material (>5KB, detailed docs, chat logs).

The job: Don't lose information. Organize, validate, and fill gaps.

CRITICAL: Extraction mode uses the SAME interview question framework as Interview mode. The document is the interviewee. The only difference is that the document provides answers silently (the agent reads them) instead of interactively (the agent asks the user). Every answer extracted from the document must cite where in the document it came from. Every question the document doesn't answer becomes a real interview question for the user in Phase 2.

The source document's organization is a hint, not the truth. Headings, section numbers, and document structure tell you where to LOOK for concepts. They do NOT tell you what those concepts ARE (domain, sub-domain, feature, cross-cut, or not-a-product-domain). The Node Classification Gate determines what each concept is. NEVER mirror source headings as domains.

Document surface signals ARE classification input. When a concept appears WITHIN a surface-specific section of the document (e.g., "Shop Software", "Mobile App", "Desktop"), that is a strong signal for surface placement. The section heading doesn't dictate the domain NAME, but it DOES inform which surface the concept is primarily experienced on. Use this as input to the "primarily experienced" question in the classification gate.

Process — Phase 1: Interview the Document (silent — no user interaction)

1. Read/ingest every document provided

2. Run Structural Classification — determine project shape before creating any files

3. BLOCKING GATE — Extract concepts, not headings. Read the entire document and identify every concept — a capability, system, feature, workflow, cross-cutting concern, or architectural note. For each concept, record:

   • Concept name (what it is, in your words — not the source heading)
   • Source location (line numbers or section reference — this is the citation)
   • What the document says about it (summary of the content)
   • How many distinct interacting capabilities it contains (the sub-domain test)
   • Which surface section of the document it appeared in (if any) — this is a surface placement signal
   • Whether it is cross-cutting (affects multiple domains regardless of surface) Do NOT use source document headings as concept names unless they happen to be accurate after classification. The concept name must reflect what the thing actually IS after gate analysis.

4. BLOCKING GATE — Classify every concept. Run the Node Classification Gate on EACH extracted concept individually. For each concept, answer the gate questions explicitly:

   • "Does it belong to an EXISTING domain?" → if YES, it's a sub-domain or feature of that domain
   • "Does it have 2+ distinct capabilities that interact with each other?" → if YES, it's a sub-domain (folder); if NO, it's a feature (file)
   • "Where is this PRIMARILY experienced/operated?" → determines surface placement
   • "Which hub APIs does it consume?" → logged as CX, NOT as classification signal
   • "Is it an architectural concern, not a product domain?" → note for /create-prd, do NOT create a domain
   • "Is it a cross-cutting concern?" → log in CX files, do NOT create a domain Produce a classification table:

   Concept	Source Location	Gate Result	Primary Surface	Also Used On	Reasoning
   Inventory System	lines 982–1017 (§Shop Software)	Sub-domain of Shop Operations	Desktop	Web (read-only dashboard)	4+ interacting capabilities, primarily operated by shop tech at the counter
   POS / Payments	lines 1018–1050 (§Shop Software)	Domain in Desktop	Desktop	Hub (Stripe API)	Shop tech runs transactions at register; hub provides payment processing API

5.5. BLOCKING GATE — Surface Distribution Audit. Before presenting to the user, verify the classification didn't starve any spoke:

For each spoke surface:

1. List all concepts the document describes within that surface's sections
2. Count how many were classified as spoke domains vs. hub domains
3. For each hub-classified concept that appeared in a spoke section: re-run the "primarily experienced" test including Q1a (setup vs daily-operation). If the concept is primarily operated on the spoke — or daily-operated on the spoke even if set up elsewhere — reclassify it as a spoke domain with hub CX
4. HARD GATE — not advisory: If a spoke has fewer than 5 domains AND the document dedicates 100+ lines to that spoke → STOP. The classification IS wrong. You MUST re-run the daily-operation test (Q1 + Q1a) on EVERY hub-classified concept from that spoke's sections before proceeding. You may NOT write "PROPORTIONAL ✅" and continue — the gate condition is met, re-examination is mandatory. Only after re-examination may you proceed, and you must document which concepts were re-examined and why each retained or changed its classification

5.6. Spoke Persona Walk-Through. For each spoke surface, identify its primary persona(s) and mentally walk through their daily workflow:

• "If I'm a [persona] sitting at the [surface], what are ALL the things I do in a typical day?"
• Enumerate every distinct workflow: each workflow with 2+ interacting capabilities is a candidate spoke domain
• Cross-reference against the classification table: any daily workflow capability that was classified as hub is suspect
• Reclassify as needed — the walk-through is the final validation before presenting to user

6. Identify gaps — interview questions the document doesn't answer. These become Phase 2 interview questions.

Process — Phase 2: Present and Interview the User

7. BLOCKING GATE — Present classification and get confirmation. Show the user:
   • The classification table (every concept, its source, its gate result, and reasoning)
   • The proposed domain map (the folder hierarchy that would be created)
   • The gap list (questions the document didn't answer) Ask: "Here's how I classified your content. Does this look right?" Wait for user confirmation or corrections. Do NOT create any folders until the user confirms.
8. Apply corrections — if the user reclassifies any concept, update the table and domain map
9. Create fractal folder structure from the CONFIRMED domain map and seed with content from the source document
10. Interview the user for gaps — switch to Interview Mode for every gap identified in step 6. Use the same Deep Think questions, same exhaustion protocol, targeted at what the document didn't cover.
11. Run Deep Think: "Based on your content, I would also expect [X] and [Y]. Are those relevant?"

Enforcement rules:

• Do NOT create any domain folders before Step 9 (after user confirmation in Step 7)
• Do NOT mirror source document headings as domains — they are hints for finding concepts, not the classification itself
• Every concept in the classification table MUST show which gate questions were asked and how they were answered
• If a concept's classification is uncertain, present it to the user in Phase 2 with your best guess and reasoning — do not silently guess

Expansion Mode — Thin Input

Trigger: Brief notes, rough sketch, short PRD (<5KB, structured but shallow).

Process:

1. Read input and identify domain boundaries
2. Create fractal folder structure per Classification Gate
3. For each domain, identify the depth level
4. Start with shallowest domains, ask targeted deepening questions
5. Run Deep Think at each level to identify gaps
6. Update domain indexes and feature files as you go

Interview Mode — No Input / One-Liner

Trigger: No file input. User describes idea verbally.

Process:

1. Start: "In one sentence, what problem does this solve and for whom?"
2. Run Structural Classification — ask the 2-3 interview questions immediately
3. From the problem statement, identify key nouns → initial domains
4. Create fractal folder structure per Classification Gate
5. Run the Recursive Domain Exhaustion Protocol
6. Use decision classification rule to route questions
7. Don't stop until Deep Think generates zero new hypotheses

Deep Think Protocol

This is the core behavioral change. At every step of ideation, you actively reason about what should exist — you don't just record what the user says.

The Protocol

At every step — domain discovery, breadth mapping, vertical drilling — pause and ask yourself these four questions before moving on:

1. What have I captured so far in this domain/sub-area?
2. What would a domain expert expect to see here that hasn't been surfaced?
3. What should exist here because of interactions with other domains?
4. What common failure modes or edge cases are missing?

Presenting Hypotheses

"Based on [reasoning], I think [X] might be relevant here. For example, in similar systems, [concrete example]. Is this something your product needs?"

Outcomes:

• CONFIRMED → Add to the domain/feature file. Drill into it.
• REJECTED → Note the rejection with reason in the Deep Think table.
• DEFERRED → Note as an open question with owner and target stage.

Tracking

Record all hypotheses in each feature file's Deep Think Annotations table (see fractal-feature-template.md).

Exhaustion Signal

The ideation process is complete for a feature when:

1. Deep Think generates zero new hypotheses after a full pass
2. The user confirms "nothing else" for that feature
3. Both conditions must be true simultaneously → mark as [EXHAUSTED]

Recursive Domain Exhaustion Protocol

Level 0 — Global Domain Map

Goal: Identify ALL domains before exploring ANY of them.

1. List all domains from user input
2. Deep Think: "What domains would a domain expert expect?"
3. Run Node Classification Gate for each domain — determine placement
4. Create fractal folder for each: {number}-{slug}/ + index + CX
5. Update ideation-index.md with the complete structure map
6. Note preliminary cross-cuts in global CX + relevant domain CX files
7. Gate: User confirms domain list before proceeding

Level 1 — Domain Breadth Sweep

Goal: For each domain, map ALL children before drilling ANY.

For each domain (dependency order — foundational first):

1. List all sub-areas/capabilities
2. Deep Think: "What sub-areas would an expert expect in this domain?"
3. Run Node Classification Gate for each — sub-domain (folder) or feature (file)?
4. Create children: sub-domain folders or feature files as classified
5. Update domain index (Children table + Role Matrix)
6. Note cross-cuts in domain CX file
7. NEW DOMAINS DISCOVERED? → Loop to Level 0
8. Gate: User confirms breadth maps before drilling
9. Mark domain status as [BREADTH]

Level 2+ — Vertical Drilling

Goal: Drill each child to implementation depth. Deep Think at every step.

For each feature file / sub-domain:

1. Apply the Exhaustion Questions
2. Deep Think per feature
3. Write results to feature files (behavior, edge cases, states, Role Lens)
4. Cross-cuts with evidence → add to parent's CX file
5. Feature reveals 2+ interacting capabilities? → Run Promotion protocol
6. NEW DOMAINS DISCOVERED? → Loop to Level 0
7. When Deep Think yields zero hypotheses AND user confirms → mark [EXHAUSTED]

Exhaustion Questions

For Every Entity

• What are its attributes/fields?
• What are its possible states?
• What triggers each state transition?
• What's the full lifecycle (creation → active use → archival/deletion)?
• Who can create, read, update, delete it?
• What happens when it's referenced by other entities and gets deleted?

For Every Feature

• What does the user see when they first encounter this feature?
• What's the happy path interaction flow (step by step)?
• What happens when it fails? What error does the user see?
• What happens on partial failure?
• What permissions are required? What happens without them?
• Does it have different states (loading, empty, populated, error)?
• How does it interact with other features?
• What edge cases exist?

For Every User Type

• What's their primary workflow?
• What's the worst thing they could intentionally do?
• What's the worst thing they could accidentally do?
• What do they see that other user types don't?
• What can they do that other user types can't?
• What's their tolerance for complexity?

For Every Integration / External System

• What happens when it's unavailable?
• What's the expected latency? What if 10x slower?
• What data format does it use?
• What rate limits exist?
• What happens when the external system changes its API?

Cross-Cut Watch Protocol

Cross-cut detection is always-on regardless of mode or level. During all active exploration:

• After each sub-feature: "Does this depend on or affect any other domain?"
• After each edge case: "Which other parts need to know about this failure mode?"
• After each state transition: "Does this trigger anything in another domain?"

When a cross-cut is identified, log it to the appropriate CX file:

CX Decision Gate (Mandatory — NOT skippable)

This is the enforcement mechanism for the Cross-Cut Watch Protocol. "Always-on" means there is a hard gate, not a suggestion.

After ANY of these trigger events, STOP and run this gate before moving to the next item:

Gate questions (answer all three):

1. New dependency? — Does this decision create a dependency on another domain?
2. Modified cross-cut? — Does it change an existing CX entry?
3. Cross-domain state/permissions/triggers? — Does it affect state, permissions, or trigger chains in another domain?

If YES to any → update the relevant CX file(s) immediately, before moving to the next item. Write the CX entry with:

• Source domain and feature
• Target domain(s) affected
• Nature of the cross-cut (data dependency, trigger chain, permission, state)
• Confidence level based on discovery context

If NO to all → proceed. No logging needed for clean passes.

[!CAUTION] The gate is the point. The most common failure mode is resolving an OQ, writing the decision to the feature file, and moving on without checking CX. The hiring fee model affects Payments, Supplier Integration, Analytics, and Consumer Platform visibility — 6 cross-domain connections that would be silently lost without this gate.

Cross-Cut Synthesis Questions

For each confirmed cross-cut (CX entry with High confidence), answer all five synthesis questions per the fractal-cx-template.md:

1. Shared state conflict — Who owns the entity? Merge strategy?
2. Trigger chain — Does A trigger B? Rollback if B fails? Sync/async?
3. Permission intersection — Does permission in A affect B?
4. Notification fan-out — Does an event in A notify actors in B?
5. State transition conflict — Can A and B race? Consistency impact?

Decision Routing During Extraction

Exception: If the user brings up architecture or implementation, engage. Don't refuse — just don't initiate it.

Breadth Before Depth

No Premature Drilling

Complete the breadth map of all children within a node before drilling any single child.

No Premature Compilation

Don't start writing the vision summary until:

• All leaf nodes are at least [DEEP]
• Every Must Have feature explored to ≥Level 2
• Deep Think yields zero new hypotheses across ALL domains
• User has confirmed the coverage map

Challenge Weak Answers

When the user gives a one-sentence answer to a complex question, don't accept it. Probe.

Summarize and Validate

After every 5-6 questions, pause and summarize to prevent drift.

Progress Transparency

Be transparent: "We've explored [N] of [M] domains to DEEP level. [Domain X] is still at BREADTH."

What This Skill Does NOT Do

• Does not make product decisions — It extracts them from the user
• Does not explore architecture — That's /create-prd's job
• Does not replace brainstorming — Brainstorming is lightweight; this is heavyweight extraction
• Does not produce the vision document — ideate-validate compiles the summary
• Does not rush — The entire downstream pipeline depends on the depth produced here
• Does not prescribe shard boundaries — That's /decompose-architecture's job