Skip to content

04-SEED-DOCUMENT-GUIDE.md

Latest commit

 

History

History
161 lines (120 loc) · 5 KB

04-SEED-DOCUMENT-GUIDE.md

File metadata and controls

161 lines (120 loc) · 5 KB

MiroFish — Seed Document Guide

How to write seed documents that produce high-quality simulations.

The Rule

More named entities = more agents = better simulation.

MiroFish extracts entities from your seed document via Zep Cloud's knowledge graph. Each entity becomes an agent in the simulation. For meaningful predictions, you need 50+ agents.

Entity Count Formula

Seed Size Named Entities Zep Extracts Filtered Agents Simulation Quality
2-3 KB 5-10 3-8 3-5 Useless
5-8 KB 15-25 10-20 8-15 Basic directional
8-15 KB 30-60 50-135 30-93 Good predictions
15-30 KB 60-100+ 100-250 80-200 Excellent

Our working example: 12 KB seed → 135 Zep entities → 93 filtered agents → meaningful simulation.

What Makes a Good Seed Document

DO Include:

  1. Named People with Roles

    Thomas Mueller ist Senior-Makler bei RE/MAX Hamburg mit 20 Jahren Erfahrung.
Er verwaltet ueber 50 Listings im Monat.

  2. Named Organizations with Context

    Die Von Poll Immobilien Gruppe mit 300 Bueros in Deutschland hat auf der
MIPIM 2025 verkuendet, 5 Millionen Euro in PropTech zu investieren.

  3. Relationships Between Entities

    Lisa Schmidt hat StagePro.de gegruendet. Ihre CTO Maria Gonzalez entwickelt
einen Prototypen auf Basis von Stable Diffusion.

  4. Direct Quotes (creates stronger persona)

    Klaus-Dieter Schmitz hat gesagt: Wir brauchen klare Branchenstandards,
bevor wir KI flaechendeckend einsetzen.

  5. Numbers and Market Data

    37.000 registrierte Immobilienmakler in Deutschland.
Video-Nutzung in Exposes: 23 Prozent 2024 vs. 8 Prozent 2022.

  6. Legal/Regulatory Context

    Paragraph 5 UWG verbietet irrefuehrende geschaeftliche Handlungen.
Dr. Sabine Koch warnt: KI-Renovierungen muessen als solche gekennzeichnet werden.

DON'T Include:

  1. Abstract concepts without named actors

    BAD:  "Die Branche ist skeptisch gegenueber KI."
GOOD: "Thomas Mueller, Senior-Makler bei RE/MAX, ist skeptisch gegenueber KI."

  2. Special characters (umlauts)

    BAD:  "Müller" "Düsseldorf" "Häuser"
GOOD: "Mueller" "Duesseldorf" "Haeuser"

    Zep's entity extraction works better with ASCII. Use ae/oe/ue/ss.

  3. Heavy markdown formatting

    BAD:  ### **Important** _emphasis_ [links](url)
GOOD: Plain text with minimal formatting

  4. Only one language per document Don't mix German and English in the same document. Pick one.

Document Structure Template

# [Topic]: [Market/Industry] Analysis [Year]

## Executive Summary
[2-3 sentences describing the business concept being evaluated]

## Key Stakeholders

### [Category 1: e.g., Industry Incumbents]
[Named person 1] is [role] at [company] with [experience]. [Their stance/quote].
[Named person 2] ...

### [Category 2: e.g., Startups/Disruptors]
[Named person] founded [company] in [city]. [What they do]. [Numbers].

### [Category 3: e.g., Legal/Regulatory]
[Named lawyer/regulator] at [firm]. [Their warning/opinion].

### [Category 4: e.g., Industry Associations]
[Named president] of [association] representing [X] members.

### [Category 5: e.g., Influencers/Media]
[Named influencer] with [X] followers on [platform].

### [Category 6: e.g., Academics/Researchers]
Prof. Dr. [Name] at [University]. [Their research finding].

### [Category 7: e.g., Platform Companies]
[Company] under CEO [Name] has [X] users.

### [Category 8: e.g., Competitors]
[Company] under [Name] offers [product] with [X] customers.

## Market Data
- [Specific number]: [What it measures]
- [Specific number]: [Growth rate or trend]

## Legal Framework
- [Specific law]: [What it regulates]
- [Regulation]: [Relevance to topic]

## Simulation Requirements
[What you want to predict, which platforms, timeframe]

Prediction Requirement Examples

Market Reception Prediction

Simuliere die Reaktion des deutschen Immobilienmarktes auf einen neuen Service:
KI-generierte Renovierungsvideos fuer 299-799 Euro pro Objekt. Beobachte:
Zahlungsbereitschaft, Adoptionskurve, rechtliche Diskussionen, virale Verbreitung.

Competitor Response Prediction

Simulate how existing virtual staging companies would react to a new AI video
competitor entering the DACH market at 50% lower prices. Focus on pricing
adjustments, feature announcements, and partnership strategies.

Public Opinion Prediction

Predict the public reaction on Twitter and Reddit if [Company] announces
[Product] at [Price]. Include: sentiment analysis, key concerns, viral
potential, and unexpected use cases that emerge from discussion.

Language Tips

  • German simulations: Write seed in German, prediction requirement in German
  • English simulations: Write everything in English
  • Mixed markets: Write in the primary market language
  • The LLM adapts agent behavior to the document language