mirofish

name	mirofish
description	MiroFish swarm intelligence — dual-mode: (1) web research swarm with parallel WebSearch, (2) prediction/forecasting engine with multi-agent perspective simulation, scenario modeling, and probability-weighted outcomes.
triggers	["research","deep dive","investigate","comprehensive search","thorough research","swarm search","mirofish","predict","forecast","what if","scenario analysis","what would happen","future prediction","simulate outcome"]
linksTo	["context-engineering","research","ai-agents","rag","sequential-thinking","make-decision"]

MiroFish — Swarm Intelligence Engine

Dual-mode: Research Swarm + Prediction Engine. Inspired by MiroFish multi-agent simulation.

Mode Detection

User says	Mode
"research X", "investigate", "deep dive", "what do we know about"	Research
"predict", "forecast", "what if", "what would happen if", "scenario", "simulate"	Predict
"mirofish" (no qualifier)	Ask which mode

---

MODE 1: Research Swarm

Question → Query Generation → Parallel Search → Deep Fetch → Synthesis → Report

Research Protocol

Step 1: Parse Research Question

Identify core topic, scope, time sensitivity, domain.

Step 2: Generate 5-8 Diverse Search Queries

Angle	Example
Factual	"[topic] statistics data 2024 2025"
Contrarian	"[topic] problems failures criticism"
Data-driven	"[topic] ROI metrics studies"
Historical	"history of [topic] evolution timeline"
Expert opinion	"[topic] expert analysis leaders opinions"
Geographic	"[topic] by country region comparison"
Regulatory	"[topic] laws regulations policy"
Future-looking	"[topic] predictions trends forecast 2026"

Step 3: Execute ALL Searches in Parallel

Use WebSearch for ALL queries simultaneously in a single message. This is the swarm.

Step 4: Deep Fetch Top 3-5 Sources

WebFetch authoritative URLs in parallel. Priority: academic > industry reports > government > news.

Step 5: Synthesize Report

## Research Report: [Topic]
### Consensus View — [cited]
### Divergent Views — [cited]
### Key Data Points — [stat + source]
### Timeline
### Confidence: High / Medium / Low / Gaps
### Sources: [numbered, linked]

Step 6: Save to Memory

Category: "research", Importance: 7, Tags: ["#research", "#mirofish"]

---

MODE 2: Prediction / Forecasting Engine

Question → Data Gathering → Variable Mapping → Agent Perspectives → Scenario Modeling → Prediction Report
    │            │                │                    │                    │                  │
 "what if    WebSearch for     identify key       6-8 archetypes     3 scenarios        weighted
  X happens"  baseline data    variables &        each reasons       (base/bull/bear)   consensus
                               causal links       independently      with probabilities

Prediction Protocol

Step 1: Frame the Prediction Question

Decompose into:

• Subject: What entity/system is being predicted?
• Intervention: What change/event triggers the prediction?
• Outcome variable: What specifically are we predicting?
• Time horizon: When? (weeks, months, years)
• Baseline: What is the current state?

Example: "What happens to tech stocks if Fed raises rates 50bps?" → Subject: tech stocks (NASDAQ composite) → Intervention: Fed funds rate +50bps → Outcome: price movement (%, direction) → Time horizon: 6 months → Baseline: current rate, current NASDAQ level

Step 2: Data Gathering (Research Swarm)

Run 5-8 parallel WebSearches focused on:

• Current state of the subject
• Historical precedents for the intervention
• Expert forecasts on related topics
• Contrarian/minority views
• Second-order effects data

This feeds real data into the prediction — not hallucinated priors.

Step 3: Variable Mapping

Identify the causal chain from intervention → outcome:

[Intervention]
    ↓
[First-order effects] — direct, immediate
    ↓
[Second-order effects] — indirect, delayed
    ↓
[Feedback loops] — self-reinforcing or dampening
    ↓
[Outcome variable]

For each link, note:

• Direction (positive/negative)
• Magnitude (strong/moderate/weak)
• Confidence (data-backed or speculative)
• Time lag

Step 4: Multi-Agent Perspective Simulation

Reason through the prediction from 6-8 distinct archetypes, each with different priors, biases, and information weighting. Each agent MUST produce a specific prediction (number or range) with reasoning.

Agent	Perspective	Weighs heavily
The Historian	"This has happened before"	Historical precedent, base rates, mean reversion
The Optimist	"Markets/systems adapt"	Innovation, resilience, positive adaptation
The Pessimist	"Tail risks are underpriced"	Downside scenarios, fragility, cascading failures
The Contrarian	"Consensus is wrong because..."	Where the crowd might be systematically biased
The Quant	"The numbers say..."	Data, correlations, statistical relationships
The Insider	"Having seen this from inside..."	Domain expertise, operational knowledge
The Systems Thinker	"Second-order effects matter most"	Feedback loops, emergent behavior, non-linear dynamics
The Geopolitical Analyst	"External forces dominate"	Regulatory, political, macro forces

For each agent, produce:

**[Agent Name]**: [Specific prediction with number/range]
Reasoning: [2-3 sentences]
Key assumption: [What must be true for this to hold]
Confidence: [0.0-1.0]

Step 5: Scenario Modeling

Synthesize agents into 3 weighted scenarios:

Bear Case (probability: X%)

• What happens: [specific outcome]
• Key drivers: [which agents' logic dominates]
• Trigger conditions: [what would cause this]

Base Case (probability: X%)

• What happens: [specific outcome]
• Key drivers: [weighted consensus]
• Assumptions: [what must hold]

Bull Case (probability: X%)

• What happens: [specific outcome]
• Key drivers: [which agents' logic dominates]
• Trigger conditions: [what would cause this]

Probabilities MUST sum to ~100%. Assign based on:

• Number of agents aligned with each scenario
• Strength of historical precedent
• Quality of supporting data

Step 6: Prediction Report

## Prediction Report: [Question]

### Question
[Full prediction question with time horizon]

### Baseline
[Current state with data, sourced]

### Causal Chain
[Intervention] → [Effect 1] → [Effect 2] → [Outcome]
Confidence in chain: X/10

### Agent Perspectives
| Agent | Prediction | Confidence | Key Assumption |
|-------|-----------|------------|----------------|
| Historian | ... | 0.X | ... |
| Optimist | ... | 0.X | ... |
| ... | ... | ... | ... |

### Scenarios
**Bear (X%)**: [outcome + reasoning]
**Base (X%)**: [outcome + reasoning]
**Bull (X%)**: [outcome + reasoning]

### Consensus Prediction
[Probability-weighted outcome]
Overall confidence: X/10
Time horizon: [when]

### Key Risks / What Would Change This
1. [Risk that invalidates prediction]
2. [New information to watch for]
3. [Leading indicator to monitor]

### Sources
[Numbered, linked — from data gathering phase]

Step 7: Save Prediction to Memory

Category: "prediction"
Importance: 8
Tags: ["#prediction", "#mirofish", "#[subject]", "#[timeframe]"]
Content: "[Question] → [Consensus prediction] (confidence: X/10, base case: Y%)"

Also save as a decision if it informs a real choice:

npx tsx ./memory/scripts/memory-runner.ts decision "[title]" "[prediction summary]" "[context]"

---

Quality Rules (Both Modes)

• Minimum 5 parallel searches — fewer is not a swarm
• Always cite sources — no unsourced claims in data gathering
• Include contrarian views — confirmation bias kills predictions
• Specific numbers — "stocks will decline" is not a prediction; "NASDAQ -8% to -15% over 6mo" is
• Explicit confidence — every claim and every agent gets a confidence score
• Time-bound — every prediction has a deadline for verification
• Save to memory — predictions are expensive, track them for calibration
• Acknowledge uncertainty — wide confidence intervals are honest, not weak

Anti-Patterns

• Do NOT skip the data gathering phase — predictions without data are just vibes
• Do NOT let all agents agree — if they do, you haven't made them distinct enough
• Do NOT assign equal probabilities to all scenarios — that's a cop-out
• Do NOT predict without a time horizon — "eventually" is not a forecast
• Do NOT present research mode output as prediction — they are different products