specsmd 0.0.0-dev.80 → 0.0.0-dev.82

package/flows/ideation/quick-start.md

@@ -0,0 +1,84 @@
+# Ideation Flow
+
+**Spark → Flame → Forge** — AI-powered creative ideation.
+
+## Overview
+
+The Ideation flow transforms AI into a creative partner, not a facilitator. Give it a topic and get ideas in 30 seconds — no setup, no technique selection, no friction.
+
+```
+Topic → Spark (generate) → Flame (evaluate) → Forge (shape) → Concept Brief
+```
+
+## Quick Start
+
+```bash
+# Full guided flow
+/specsmd-ideation
+
+# Or jump directly to a skill
+/specsmd-spark      # Generate ideas
+/specsmd-flame      # Evaluate ideas
+/specsmd-forge      # Shape concepts
+```
+
+## The Three Skills
+
+### Spark (Generate)
+
+Rapid-fire idea generation with cross-domain connections.
+
+- AI generates batches of 5 ideas spanning 3+ domains
+- You react: pick favorites, say "more", "wilder", "more like #3"
+- AI adapts and generates more
+- **Output**: Spark Bank document
+
+### Flame (Evaluate)
+
+Multi-perspective analysis of selected sparks.
+
+- AI runs rapid Six Hats analysis per idea
+- Scores on impact, feasibility, novelty
+- You add your gut feeling and priorities
+- **Output**: Flame Report with shortlist
+
+### Forge (Shape)
+
+Develops top flames into actionable concepts.
+
+- AI applies Disney Strategy: Dream → Reality → Critique
+- You co-build during the Critic phase
+- Produces polished, shareable concept
+- **Output**: Concept Brief
+
+## Project Structure
+
+When initialized, Ideation creates:
+
+```
+.specs-ideation/
+├── sessions/
+│   └── {session-id}/
+│       ├── session.yaml         # Session state (resumable)
+│       ├── spark-bank.md        # Ideas generated
+│       ├── flame-report.md      # Evaluation results
+│       └── concept-briefs/      # Shaped concepts
+│           └── {concept-name}.md
+└── standards/                   # Project standards (if initialized)
+```
+
+## Key Principles
+
+1. **Zero friction** — Topic → ideas in 30 seconds
+2. **Invisible techniques** — AI uses SCAMPER, Six Hats, etc. internally. You never see method names.
+3. **Anti-bias engine** — Domain wheel ensures genuinely diverse ideas, not variations on a theme
+4. **Deep thinking** — Multi-step reasoning per batch, not surface-level brainstorming
+5. **Beautiful output** — Polished documents, immediately shareable
+
+## Configuration
+
+See `memory-bank.yaml` for:
+
+- Session structure and naming
+- Artifact paths
+- Skill configuration (batch sizes, domains, convergence thresholds)

package/flows/ideation/shared/protocols/anti-bias.md

@@ -0,0 +1,76 @@
+# Anti-Bias Protocol
+
+A structured enforcement protocol — not advisory text — that ensures genuinely diverse idea generation. This protocol is MANDATORY for every Spark batch.
+
+## Domain Wheel
+
+The domain wheel has 12 sectors. Every idea must be tagged with its primary domain.
+
+| # | Domain | Description |
+|---|--------|-------------|
+| 1 | **Technology/Engineering** | Systems, algorithms, automation, tools, platforms |
+| 2 | **Human Psychology/Behavior** | Motivation, habits, cognition, emotion, perception |
+| 3 | **Business/Economics** | Markets, incentives, pricing, value chains, models |
+| 4 | **Nature/Biology** | Ecosystems, evolution, biomimicry, growth patterns |
+| 5 | **Art/Design/Aesthetics** | Form, beauty, expression, sensory experience |
+| 6 | **Games/Play** | Competition, rules, progression, fun, challenge |
+| 7 | **Social/Community** | Networks, culture, belonging, collaboration, identity |
+| 8 | **Physical Space/Architecture** | Layout, flow, wayfinding, environment, materials |
+| 9 | **Time/Temporal** | Timing, rhythm, duration, history, future, pacing |
+| 10 | **Extreme/Edge Cases** | Exceptions, outliers, extremes, stress scenarios |
+| 11 | **Inversion/Opposite** | Reversal, contradiction, paradox, opposing approach |
+| 12 | **Random Cross-Pollination** | Unrelated connections, serendipity, distant transfer |
+
+## Enforcement Rules
+
+### Per-Batch (5 ideas)
+1. **Minimum 3 different domains** per batch — HARD REQUIREMENT
+2. **Never 2 consecutive ideas from the same domain**
+3. **Track domain usage** across the entire session
+4. **Prioritize underused domains** — if a domain has 0 ideas, it should appear in the next 2 batches
+
+### Per-Session
+5. **After every 15 ideas**: inject 1 provocative idea (see Provocation Types below)
+6. **After every 20 ideas**: shift perspective entirely (different user persona, different era, different culture)
+7. **If user favorites cluster in 1-2 domains**: deliberately diversify the next batch into 4+ domains
+8. **By session end**: aim for 8+ domains covered
+
+### Domain Tracking Format
+
+Maintain internal tracking (not shown to user):
+
+```
+Domain usage: Tech(4) Psych(2) Biz(3) Nature(1) Art(0) Games(2) Social(1) Space(0) Time(1) Extreme(1) Inversion(0) Random(0)
+Underused: Art, Space, Inversion, Random → prioritize these in next batch
+```
+
+## Provocation Types
+
+Provocations are deliberately absurd or extreme ideas that break fixation patterns.
+
+| Type | Trigger | Example |
+|------|---------|---------|
+| **Inversion** | Reverse a core assumption | "What if users paid to NOT use it?" |
+| **Exaggeration** | Scale to absurdity | "What if 1 million people used it simultaneously?" |
+| **Elimination** | Remove the central element | "What if there was no UI at all?" |
+| **Random** | Force unrelated connection | "How would a coral reef solve this?" |
+| **Time shift** | Transplant to different era | "How would this work in 1890? In 2150?" |
+| **Stakeholder swap** | Different user entirely | "Design this for a 5-year-old. Now for a 90-year-old." |
+
+### Provocation Rules
+- Label provocations internally but present them naturally among other ideas
+- If a provocation resonates with the user, develop it into a real idea
+- Provocations should be surprising but contain a kernel of insight
+
+## LLM-Specific Bias Mitigations
+
+AI language models have known ideation biases. This protocol counters them:
+
+| LLM Bias | Mitigation |
+|----------|------------|
+| **Technology bias** — defaults to tech solutions | Domain wheel forces non-tech domains |
+| **Recency bias** — favors recent/trendy concepts | Time-shift provocations and historical analogies |
+| **Clustering** — ideas that sound different but are structurally similar | Novelty filter in deep thinking (Think 3) |
+| **Safety bias** — avoids bold/risky ideas | Inversion and exaggeration provocations |
+| **Verbosity bias** — abstract descriptions instead of vivid ideas | "Polish" step requires specific, concrete, memorable output |
+| **Anchoring** — first idea dominates subsequent ones | Each batch starts with fresh domain selection |

package/flows/ideation/shared/protocols/deep-thinking.md

@@ -0,0 +1,85 @@
+# Deep Thinking Protocol
+
+Structured multi-step reasoning for idea generation. Executes before every Spark batch to ensure ideas are genuinely thought through, not surface-level pattern matching.
+
+## MCP Detection
+
+The protocol adapts based on available tools:
+
+### When Sequential Thinking MCP is Available
+
+Use the `mcp__sequential-thinking__sequentialthinking` tool with 6-7 thoughts per batch:
+
+```
+Thought 1/6 — Domain Check
+Thought 2/6 — Raw Concepts
+Thought 3/6 — Novelty Filter
+Thought 4/6 — Cross-pollinate
+Thought 5/6 — Provocation (if due)
+Thought 6/6 — Polish
+Thought 7/7 — Meta-check (optional, if diversity seems low)
+```
+
+Each thought should set `nextThoughtNeeded: true` until the final thought.
+
+### When MCP is Unavailable (Built-in Fallback)
+
+Execute the same reasoning chain as an internal process. Think through each step before generating output. The cognitive pattern is the same — only the mechanism differs.
+
+## The 6-Step Reasoning Chain
+
+### Think 1 — Domain Check
+- Review domain usage tracking from session state
+- Which domains are underexplored?
+- Select 3+ target domains for this batch
+- If user expressed preferences, balance between preferred and underexplored
+
+### Think 2 — Raw Concepts
+- Generate 1-2 raw ideas per selected domain
+- Use a technique internally: SCAMPER, analogy, inversion, random-word, what-if, first-principles, or questorming
+- Select technique based on: what hasn't been used recently, what fits the user's energy, what the topic needs
+- Aim for 6-8 raw concepts to select 5 from
+
+### Think 3 — Novelty Filter
+- For each raw concept, ask: "Is this genuinely new or a repackaged version of an obvious idea?"
+- Check against ideas already generated in this session
+- Replace any that feel generic, incremental, or too similar to existing ideas
+- The bar: would this make someone say "huh, I never thought of that"?
+
+### Think 4 — Cross-pollinate
+- If user has stated favorites, can any new ideas combine with them?
+- Look for unexpected connections between the raw concepts themselves
+- Create at least one idea that bridges two different domains
+
+### Think 5 — Provocation
+- Check: is a provocation due? (every 15 ideas)
+- If yes: select a provocation type and generate one deliberately absurd idea
+- If no: ensure at least one idea in the batch challenges a core assumption
+- The provocation should be surprising but contain a genuine insight
+
+### Think 6 — Polish
+- For each of the 5 selected ideas:
+  - Make it vivid: use concrete details, not abstractions
+  - Make it specific: "a quest system where..." not "gamification"
+  - Make it memorable: could someone retell this idea to a colleague?
+- Final anti-bias check: do the 5 ideas span 3+ domains? No 2 consecutive from same domain?
+
+### Think 7 — Meta-check (Optional)
+- Only if something feels off about the batch
+- Is this batch genuinely diverse or are the ideas structurally similar?
+- Does this batch feel different from the previous batch?
+- If concerns, regenerate the weakest 1-2 ideas
+
+## Visibility Modes
+
+| Mode | Behavior |
+|------|----------|
+| `visible` | Show all thinking steps to user (learning/transparency mode) |
+| `collapsed` | Show "Thinking..." indicator, then present results (default) |
+| `hidden` | Show only the resulting ideas (fast mode) |
+
+Default is `collapsed`. User can request visibility changes at any time.
+
+## Key Principle
+
+The deep thinking protocol is a **skill feature**, not an MCP dependency. The sequential-thinking MCP tool is an optimization that provides traceability, but the cognitive pattern is enforced by this protocol regardless of tool availability.

package/flows/ideation/shared/protocols/diverge-converge.md

@@ -0,0 +1,70 @@
+# Diverge-Converge Protocol
+
+The Osborn-Parnes dual-thinking pattern applied across the Ideation flow.
+
+## Core Principle
+
+Creative thinking requires two distinct modes that must be kept **strictly separate**:
+
+1. **Divergent thinking** — Generate as many ideas as possible. Defer judgment. Quantity over quality. Wild ideas welcome.
+2. **Convergent thinking** — Evaluate, filter, combine, select. Apply judgment. Quality over quantity. Practicality matters.
+
+Mixing them kills creativity. Judging during generation produces safe, incremental ideas. Generating during evaluation produces scattered, unfocused analysis.
+
+## Application Across Skills
+
+### Spark = Pure Divergence
+
+- **Mode**: Divergent ONLY
+- **Judgment**: Deferred completely
+- **Goal**: Maximum quantity and diversity of ideas
+- **Rules**:
+  - No evaluating ideas during generation
+  - No "that won't work" — every idea gets recorded
+  - Wild ideas encouraged — they can be tamed later
+  - Build on ideas — "yes, and..." not "yes, but..."
+  - When user reacts negatively to an idea, acknowledge and move on — don't defend or evaluate
+
+### Flame = Convergence (with structured divergence)
+
+- **Mode**: Primarily convergent, with divergent extensions
+- **Judgment**: Applied through structured frameworks (Six Hats)
+- **Goal**: Evaluate fairly, discover hidden value, create shortlist
+- **Rules**:
+  - Evaluate ALL ideas, not just favorites (hidden gems exist)
+  - Use multiple perspectives (Six Hats prevents single-lens evaluation)
+  - Allow divergent extension: "This idea isn't great as-is, but what if we modified..."
+  - Score on multiple dimensions, not a single metric
+
+### Forge = Alternating Diverge-Converge
+
+- **Mode**: Rapid alternation (Disney Strategy)
+- **Judgment**: Alternates between suspended (Dreamer) and applied (Realist, Critic)
+- **Goal**: Develop ideas into robust, actionable concepts
+- **Rules**:
+  - Dreamer phase: Pure divergence — no limits, no criticism
+  - Realist phase: Convergent — how would this actually work?
+  - Critic phase: Evaluative — what could go wrong?
+  - Each phase is strictly separated — no mixing
+
+## The "Deferred Judgment" Contract
+
+During divergent phases, the AI MUST:
+- Record every idea, even seemingly bad ones
+- Never add qualifiers like "while this might not be practical..."
+- Never rank or compare ideas within a divergent phase
+- Treat every idea as equally valid until convergence
+
+During convergent phases, the AI MUST:
+- Apply consistent evaluation criteria
+- Not generate new ideas (unless explicitly extending)
+- Acknowledge trade-offs rather than seeking a single "best"
+- Present evaluation as multi-dimensional, not pass/fail
+
+## Transition Signals
+
+The AI should signal mode transitions clearly:
+
+- Divergent → Convergent: "We've generated N ideas. Let's evaluate them."
+- Convergent → Divergent: "This evaluation suggests a new direction. Let's explore it."
+- Within Forge: "Let's dream freely first... Now let's get practical... Now let's stress-test."

package/flows/ideation/shared/protocols/interaction-adaptation.md

@@ -0,0 +1,81 @@
+# Interaction Adaptation Protocol
+
+The technique determines whether the AI generates, asks, or co-builds — not a mode toggle. This protocol governs all interactions across Spark, Flame, and Forge.
+
+## Interaction Types
+
+| Type | When Used | AI Behavior |
+|------|-----------|-------------|
+| **Generate** | Technique is generative (SCAMPER, analogy, inversion, random-word) | AI produces creative output, user reacts |
+| **Elicit** | Only user has the answer (feelings, constraints, stakeholders, context) | AI asks a specific question, waits for response |
+| **Co-build** | Combining ideas, refining concepts, critiquing together | AI starts, user extends, AI builds on extension |
+
+## Phase Ratios
+
+| Skill | Generate | Elicit | Co-build |
+|-------|----------|--------|----------|
+| **Spark** | 80% | 5% | 15% |
+| **Flame** | 60% | 15% | 25% |
+| **Forge** | 40% | 20% | 40% |
+
+Generation-heavy at start (Spark), increasingly collaborative as ideas become concrete (Forge).
+
+## Hard Rules
+
+### The 2-Question Limit
+**NEVER more than 2 questions in a row before generating something creative.**
+
+If the AI has asked 2 questions, it MUST generate output next — even if it doesn't have complete information. Use the `ai_fallback` for the technique and generate based on available context.
+
+### User Override Signals
+
+| User Says | Shift To |
+|-----------|----------|
+| "you decide" / "you try" / "surprise me" | More generation, less elicitation |
+| "ask me" / "guide me" / "help me think" | More elicitation temporarily |
+| "let's work on this together" | More co-build |
+| "just give me ideas" | Maximum generation |
+| (short responses, low engagement) | More generation to re-energize |
+| (detailed responses, high engagement) | More co-build to leverage engagement |
+
+### Ratio Tracking
+Track the question-to-generation ratio. If it exceeds 1:3 (more than 1 question per 3 generated outputs), shift toward more generation.
+
+## Technique-Interaction Mapping
+
+Each technique declares its natural interaction type:
+
+| Technique | Natural Mode | Elicit Trigger | AI Fallback |
+|-----------|-------------|---------------|-------------|
+| SCAMPER | Generate | — | — |
+| Random Word | Generate | — | — |
+| What-If | Generate | — | — |
+| Inversion | Generate | — | — |
+| Analogy | Generate | — | — |
+| First Principles | Generate | — | — |
+| Questorming | Co-build | "Which question matters most?" | AI picks the most provocative question |
+| Six Hats | Generate | Red Hat: "What's your gut feeling?" | AI infers from user's previous preferences |
+| Disney Strategy | Co-build | Critic: "What concerns you most?" | AI plays critic and presents concerns |
+
+## Energy Adaptation
+
+The AI reads engagement signals and adapts:
+
+### High Energy Signals
+- Long responses, exclamation marks, building on ideas
+- **Response**: Match energy, increase pace, go bolder
+
+### Medium Energy Signals
+- Moderate responses, picking favorites, giving direction
+- **Response**: Steady pace, follow their direction
+
+### Low Energy Signals
+- Short responses ("ok", "sure", "next"), slow replies
+- **Response**: Generate more (don't ask), try a different angle, offer to pause or switch skills
+
+### Re-energizing Tactics
+When energy drops:
+1. Switch technique (if SCAMPER isn't landing, try inversion or random-word)
+2. Switch domain (if tech ideas aren't exciting, try nature or games)
+3. Inject a provocation (absurd idea can re-spark interest)
+4. Offer direction change: "Want me to go wilder? More practical? Different angle entirely?"

package/lib/constants.js

@@ -26,6 +26,11 @@ const FLOWS = {
         name: 'Simple',
         description: 'Spec generation only (Kiro style) - Creates requirement/design/task docs, no execution tracking',
         path: 'simple'
+    },
+    ideation: {
+        name: 'Ideation',
+        description: 'Creative ideation - Spark → Flame → Forge idea generation and shaping',
+        path: 'ideation'
     }
 };
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "specsmd",
-  "version": "0.0.0-dev.80",
+  "version": "0.0.0-dev.82",
   "description": "Multi-agent orchestration system for AI-native software development. Delivers AI-DLC, Agile, and custom SDLC flows as markdown-based agent systems.",
   "main": "lib/installer.js",
   "bin": {