idea-darwin（达尔文思想）

Idea Darwin Engine

A round-based idea iteration system that treats ideas as competing organisms — scoring, selecting, crossing, and evolving them through structured rounds to surface the strongest concepts.

The metaphor is Darwinian selection: ideas that score well survive and evolve; weak ideas get flagged for the user to decide on; crossbreeding produces novel offspring. This prevents the common failure mode of brainstorming where all ideas stay equally vague forever.

Commands

Arguments are extracted from $ARGUMENTS. If no arguments are provided, show current status and suggestions.

init Optional Parameters

Example: /idea-darwin init --budget 8 --actions 3 --disruption 5

These parameters are written to config.yaml and can be manually edited later. Unspecified parameters use defaults.

File Structure

The system maintains this structure in the working directory (created during initialization):

project/
├── ideas.md          # User's raw ideas (read-only — never modify this file)
├── config.yaml       # System configuration and state
├── stimuli.md        # External stimulus entries (user-editable, system read-only)
├── cards/            # Idea cards
│   ├── IDEA-0001.md
│   └── ...
├── rounds/           # Round reports
│   ├── round-000.md
│   └── ...
├── reports/          # Leaderboard and cluster reports
│   └── leaderboard.md
└── graph/            # Relationship graph
    └── relations.json

Idea Numbering

• Format: IDEA-XXXX (4-digit zero-padded, starting from 0001)
• Globally unique — numbers are never recycled, even for dormant or removed ideas
• config.yaml tracks the next available number in next_idea_id
• Each idea card records lineage via parent_ids and child_ids

Idea State Machine

seed → exploring → refining → crossing → validated → dormant

State transitions are recommended by the system based on scores and action results, but the user always has final say.

Scoring System

6 Scoring Dimensions (1–10 each)

Three-Layer Scoring Formula

The scoring is layered so that each layer captures a different strategic lens — Survival measures standalone quality, Development measures growth potential, and Priority blends both with recency and diversity bonuses to prevent the pool from stagnating.

Survival = 0.10×Novelty + 0.20×Feasibility + 0.20×Value
         + 0.20×Logic + 0.10×CrossPotential + 0.20×Verifiability

Development = 0.30×Novelty + 0.30×CrossPotential
            + 0.20×VariationPotential + 0.20×Freshness

Priority = 0.50×Survival + 0.30×Development
         + 0.10×NewIdeaBoost + 0.10×DiversityBonus

• VariationPotential (0–10): How many distinct directions can this idea still branch into. Decreases as more children are derived.
• Freshness (0–10): Has this idea changed recently? Decays when an idea goes unprocessed for several rounds.
• NewIdeaBoost: Newly pooled ideas get +2 for their first 2 rounds, then 0.
• DiversityBonus: Ideas belonging to underrepresented categories in the pool get +1–2.

When scoring, refer to the prompt templates in references/prompts.md.

Initialization Flow

When the user runs /idea-darwin init:

1. Locate ideas.md

- Check if ideas.md exists in the current directory - If not found, use AskUserQuestion to ask for the path

1. Create directories and config

- Create cards/, rounds/, reports/, graph/ - Generate config.yaml (merge user-provided optional params; use defaults for unspecified ones):

source:
  ideas_file: "./ideas.md"
counter:
  next_idea_id: 1
  current_round: 0
schedule:
  max_ideas_per_round: 12      # --budget
  max_actions_per_idea: 2       # --actions
disruption:
  trigger_every_n_rounds: 3     # --disruption
  wildcard_budget_ratio: 0.10

- Generate stimuli.md (external stimulus file in the project root) with an initial template and editing instructions

1. Parse ideas.md

- Identify each independent idea (typically separated by ## headings) - Do not modify ideas.md — it is always read-only

1. Generate a card for each idea

- Assign ID (starting from IDEA-0001) - Tag (topic / problem type / stage / style / resource tags) - Generate full card following assets/card-template.md format - Initial scoring (6 dimensions + 3-layer composite scores) - Write to cards/IDEA-XXXX.md

1. Generate round-000.md report

- List all initial card summaries, scores, and suggested next steps - Write to rounds/round-000.md

1. Present results to user

- Display a summary table of all cards (ID, title, score, suggested action) - Inform the user about stimuli.md and how to use it: > stimuli.md has been created in the project root. Every N rounds, the system randomly selects up to 5 enabled entries as external stimuli to break mental ruts. > You can edit this file anytime to add/modify/disable entries. Each entry follows this format: > `` > ## [Entry Title] > - status: enabled / disabled > - [Content description] > `` - Wait for user confirmation before entering the formal round cycle

Round Execution Flow

When the user runs /idea-darwin round, execute these 9 steps strictly in order:

Step 1: Import New Ideas

• Every round starts by checking the default idea document: read the file pointed to by source.ideas_file in config.yaml
• Compare the default ideas.md against existing cards to identify new additions
• For new ideas: tag → generate card → initial scoring
• New ideas are marked stage: seed and receive the NewIdeaBoost
• This is a mandatory pre-step for every round — the system must not loop only over existing cards while ignoring incremental additions to the source file

Step 2: Update Relationship Graph

• Based on all cards' tags and content, compute inter-idea similarity
• Identify potential cross candidates (high similarity = neighbor-cross candidates; low similarity but complementary = distant-cross candidates)
• Update graph/relations.json

Step 3: Global Score Refresh

• Recalculate the 6-dimension scores for every active idea (not dormant)
• Update Survival / Development / Priority
• Flag: high-potential (Priority > 7), stagnant (unchanged for 3+ rounds), low-potential (Priority < 3)

Step 4: Budget Allocation

Distribute slots according to max_ideas_per_round:

• 25% for newly pooled ideas
• 30% for deepening high-priority ideas
• 20% for deriving from high-potential ideas
• 15% for cross experiments
• 10% for disruption / random exploration

Select ideas by Priority Score descending. Each idea gets at most max_actions_per_idea actions.

Step 5: Decide Actions

For each selected idea, assign actions based on the state machine (see table above). When the state machine is insufficient, use score-driven rules:

• High Novelty + Low Verifiability → deepen
• High Novelty + High CrossPotential → cross
• High Value + Low Feasibility → critique + refine
• High Feasibility + Low Logic → deepen
• High duplication → merge

Step 6: Execute Actions

Process each selected idea in turn. Refer to references/actions.md for detailed execution specs, and references/prompts.md for prompt templates.

Execution flow per idea:

1. Read the full card
2. Follow the corresponding prompt template for the action type — think deeply
3. Produce structured output:

- Updated card content - New derived ideas (auto-assign IDs, update parent_ids/child_ids on both sides) - Score changes - Discovered risks - Suggested next actions

1. Write the updated card file

Newly derived ideas also require full card generation and initial scoring.

Step 7: Validate and Deduplicate

• Run two-layer validation on all newly produced ideas:

- Layer 1 (Fatal flaw check): Logical contradiction? No real demand? Completely infeasible? Near-duplicate? - Layer 2 (Value judgment): Worth continuing? What next? Biggest risk?

• Validation failure → mark validated: false, lower priority, report to user
• The user decides whether to dormant or remove — the system only recommends
• Run similarity check across the entire pool and merge near-duplicates

Step 8: Generate Round Report

Write to rounds/round-XXX.md, including:

• Ideas processed this round and actions executed
• Newly produced idea summaries
• Top 5 score changes (up/down)
• Ideas that failed validation (require user decision)
• Recommended priorities for next round
• If disruption round: which external stimuli were used, which innovation frameworks applied

Step 8.5: Output Round Briefing

After all file updates are complete, output a round briefing directly in the conversation (not written to file) so the user can understand the full picture without opening any files. Format:

## Round N Briefing

### Actions Taken
| ID | Title | Action | Key Output |
|---|---|---|---|
| IDEA-XXXX | ... | deepen/derive/cross/critique/... | One-line summary |

### Card Changes
| ID | Title | Score Change | State Change | Notes |
|---|---|---|---|---|
| IDEA-XXXX | ... | Priority X.XX→X.XX | seed→exploring | Reason for change |

### New Ideas (Detail)

For each newly produced idea this round:

#### IDEA-XXXX: [Title]
- **Source**: [cross/derive/stimulus] from IDEA-XXXX + IDEA-XXXX
- **Core Question**: [one paragraph]
- **Key Insight**: [what's most valuable about this idea]
- **Initial Score**: Priority X.XX (Survival X.XX / Development X.XX)
- **Current Tension**: [biggest unresolved question]
- **Suggested Next Step**: [deepen/derive/cross/critique + specific direction]

### Current Rankings
| Rank | ID | Title | Priority | Stage |
|---|---|---|---|---|

### Next Round Preview
- [Suggested priority actions]
- [Whether a disruption round is coming soon]

### Decisions Needed
> List key decisions the user needs to make this round. Each item **must include the idea's full title** (not just the ID) along with context and suggested options.
> Examples:
> 1. **IDEA-0002 "Learn Coding Agent Tutorial Site" overlaps with IDEA-0004 "Progressive Unboxing Pedagogy"** — Merge or reposition? Suggestion: [...]
> 2. **IDEA-0006 "Auto-Diagnostic Thread" failed validation** — Dormant, pivot, or keep observing?
> 3. **IDEA-0003 "Thermal Management Ontology" downgraded** — Accept downgrade from ontology to structured Excel as first step?
>
> If no user decisions are needed this round, write "No decisions needed this round."

This briefing is the primary interface between the user and the system — keep it complete and readable. The "Decisions Needed" section is the most important interaction point: the system should clearly list decision items with recommendations, rather than leaving the user to discover issues from reports.

Step 9: Save State

• Update all modified card files
• Increment current_round in config.yaml
• Update next_idea_id
• Update reports/leaderboard.md (Top 10 rankings)

External Stimuli and Disruption Rounds

Every trigger_every_n_rounds rounds (default 3), external stimuli are introduced in a "disruption round." The purpose is to prevent the idea pool from converging on a local optimum by forcing fresh perspectives.

External Stimulus File (stimuli.md)

stimuli.md lives in the project root. It is created during initialization and maintained by the user. Format:

## Industry Trend: Embodied AI Accelerating to Production
- status: enabled
- Embodied AI moving from lab to production line in 2026, robot operating costs down 60%

## Tech Shift: Unified Multimodal Model Architecture
- status: enabled
- Vision, language, and code unified into single models, API paradigm changing

## Contrarian View: SaaS Is Dead
- status: disabled
- Open-source + self-hosted deployments replacing traditional SaaS subscriptions

• Each entry: ## heading + status field + content description
• enabled = participates in random selection; disabled = skipped
• Users can add, edit, or toggle entries at any time
• The system only reads this file — never writes to it

Disruption Round Mechanics

1. Shuffle priorities: Give low-priority but novel ideas a chance to surface
2. Distant cross: Force crossbreeding between ideas with non-overlapping tags
3. External stimuli: Randomly select up to 5 enabled entries from stimuli.md
4. Assumption reversal: For top-scoring ideas, ask "What if the core assumption is wrong?"
5. Innovation frameworks: Apply SCAMPER / First Principles / Analogical Transfer / Constraint Reversal / Extreme User Perspective

Each round also reserves wildcard_budget_ratio (default 10%) of slots for non-optimal-path exploration.

Dormancy and Wake

/idea-darwin dormant IDEA-XXXX

1. Read the card, set stage to dormant
2. Record reason and round number in the card's "Dormancy Log"
3. The idea will no longer be selected for deepen/derive, but still participates in cross and similarity matching

/idea-darwin wake IDEA-XXXX

1. Read the card, restore stage to the pre-dormancy state (inferred from history, typically exploring)
2. Update the "Dormancy Log"

The system never auto-dormants any idea. Only the user can make that call.

/idea-darwin status

Read config.yaml and all cards, then output:

• Current round number
• Total ideas in pool (broken down by state)
• Top 10 rankings (by Priority Score descending)
• What was processed in the most recent round
• Suggested next steps

Key Constraints

1. ideas.md is read-only: Never modify the user's ideas.md
2. IDs never recycle: Numbers are never reused under any circumstance
3. Branch limit: Each idea can derive at most 2–3 children per round
4. Per-round dedup: Similarity check is mandatory before a round concludes
5. User control: The system never auto-dormants or removes ideas — only recommends
6. Structured output: Every action must produce structured results
7. Consistent card format: All cards strictly follow assets/card-template.md
8. Traceable reports: Every round report must contain a complete operation log

Reference Files

• references/actions.md: Detailed specs and use cases for each standard action
• references/prompts.md: Prompt templates for each action
• assets/card-template.md: Card file template