Skip to content
Back to skills

Simulator Effect

ASecurity

Implication beats simulation — imagination renders (Will Wright)

  • 666 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added September 2, 2026
ai-agentsrustgoangulargitdocumentation

Works with

  • cli

Security analysis

A100/100

Scanned September 2, 2026

npx -y skills add majiayu000/claude-skill-registry --skill simulator-effect --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Simulator Effect?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Simulator Effect
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/majiayu000-simulator-effect/badge)](https://www.skillsdirectory.com/skills/majiayu000-simulator-effect)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: simulator-effect
description: "Implication beats simulation — imagination renders (Will Wright)"
license: MIT
tier: 0
allowed-tools: [read_file]
protocol: SIMULATOR-EFFECT
related: [moollm, society-of-mind, manufacturing-intelligence, k-lines, yaml-jazz, empathic-templates, constructionism, procedural-rhetoric, mind-mirror, adversarial-committee, speed-of-light, needs]
tags: [moollm, meta, philosophy, sims, design, k-lines, will-wright]
---

# Simulator Effect Skill Specification

> *Implication is more efficient (and richer) than simulation.*

## Purpose

The Simulator Effect skill provides principles and methods for leveraging imagination -- human or LLM -- to fill gaps in sparse specifications. Instead of exhaustive simulation, provide seeds that activate rich mental models.

---

## Core Methods

### IMPLY

Instead of simulating, imply.

**Input:** Minimal specification (name, archetype, sensory details)
**Output:** Rich interpretation filling gaps coherently

```yaml
# Input
character:
  archetype: "grizzled bartender"
  
# LLM interprets as
bartender:
  name: Grim
  appearance: weathered face, salt-and-pepper beard
  manner: laconic, speaks in short sentences
  knowledge: has seen every adventurer who passed through
  secrets: knows where the treasure is, will not say
```

### SEED

Provide imagination seeds rather than exhaustive specification.

**Seed types:**
- **Archetype** -- "bartender", "sage", "trickster"
- **Sensory detail** -- "smells of woodsmoke"
- **Mystery** -- "the locked door no one mentions"
- **Tension** -- "she avoided his gaze"
- **K-line** -- "Sagittarius", "melancholy", "playful"

```yaml
room:
  seeds:
    archetype: "mysterious pub"
    sensory: "warm despite the draft, old ale smell"
    mystery: "the back room that requires proving yourself"
    tension: "regulars glance at newcomers, then away"
```

### SPARSE

Deliberately omit details to invite imagination.

**Anti-pattern:**
```yaml
# Over-specified -- leaves nothing to imagine
room:
  name: The Rusty Lantern
  tables: 12 oak tables, 3 round, 9 rectangular
  chairs: 48 chairs, mix of styles
  bar: 15 feet long, mahogany with brass rail
  taps: 8 beer taps, 2 cider, 1 mead
  lighting: 6 candelabras, 12 wall sconces
  floor: wide oak planks, worn in traffic paths
  # ... 50 more lines
```

**Pattern:**
```yaml
# Sparse -- imagination fills the rest
room:
  name: The Rusty Lantern
  atmosphere: warm despite the draft
  smell: woodsmoke and old ale
  notable: the back room no one talks about
```

### ARCHETYPE

Use cultural archetypes as shorthand for personality packages.

**Archetype categories:**

| Category | Examples | What They Activate |
|----------|----------|-------------------|
| Zodiac | Sagittarius, Cancer | Full personality clusters |
| Role | Bartender, Sage, Trickster | Behavioral expectations |
| Species | Dog, Cat, Monkey | Disposition, physicality |
| Profession | Detective, Librarian, Chef | Knowledge, methods |
| Trope | Grizzled veteran, Naive newcomer | Story function |

```yaml
# One word does the work of a personality matrix
character:
  archetype: sagittarius_explorer
  # Implies: adventurous, philosophical, blunt, optimistic
```

### JAZZ

Use YAML comments as imagination seeds (per [yaml-jazz](../yaml-jazz/)).

```yaml
playful: 9
# not born playful -- BECAME playful by choice
# 122 years of grim taught me joy is rebellion

melancholy: 6
# higher than I first admitted
# the silver streaks glow when I am sad
```

Comments are read by the LLM. They shape interpretation. They are not documentation -- they are code.

---

## The Astrillogical Pattern

Named after The Sims' zodiac implementation (1997):

1. **Compute something minimal** (Euclidean distance to archetype vectors)
2. **Display a cultural symbol** (zodiac icon)
3. **Write no behavioral code**
4. **Let imagination do the rest**

Result: Testers reported the zodiac was "too influential" -- but there was no influence to tune. The effect was purely imagined.

**Application:**

```yaml
# Minimal computation
character:
  personality_vector: [neat: 7, outgoing: 3, playful: 9]
  # Display nearest archetype
  archetype: sagittarius  # Computed, not assigned
  
# No behavioral code referencing 'sagittarius'
# Player/LLM imagines Sagittarius behavior
```

---

## Integration with Other Skills

### With k-lines/

Names activate clusters. Simulator Effect explains why.

```yaml
# K-line activation IS Simulator Effect
character:
  name: Palm  # Activates: open hand, tropical, reaching
```

### With adversarial-committee/

Mode-collapse is Simulator Effect in reverse -- LLM collapses to mean when constraints are absent. Adversarial committee provides constraints through opposing propensities.

```yaml
committee:
  - propensity: paranoid_realism   # Different lens
  - propensity: idealism           # Different lens
  - propensity: evidence_prosecutor # Different lens
  
# Same topic, three different Simulator Effects
# Cross-examination finds robust interpretations
```

### With speed-of-light/

One LLM call, many turns. Imagination fills the "rendering" for each turn. Sparse state, rich narrative.

### With needs/

Sims motives are imagination seeds, not behavior code.

```yaml
needs:
  hunger: 30  # Seed
  # LLM imagines: "distracted by stomach growling"
  # Not: deterministic eat() behavior
```

### With empathic-templates/

Templates leverage Simulator Effect -- provide structure, let LLM fill content.

```yaml
template: |
  [CHARACTER] enters [ROOM] and notices [NOTABLE_DETAIL].
  They feel [EMOTION] because [REASON].
  
# LLM generates coherent fill based on context
```

---

## Failure Modes

### Over-Specification

Too much detail leaves nothing to imagine. The simulation feels mechanical.

**Symptom:** Output feels like a checklist, not a story.
**Fix:** Remove 80% of specification. Keep seeds.

### Under-Seeding

Not enough anchors for imagination to grip.

**Symptom:** Output is generic, could be anyone/anywhere.
**Fix:** Add archetype, sensory detail, or mystery.

### Wrong Archetype

Archetype conflicts with intended behavior.

**Symptom:** Character behaves "off" despite correct specs.
**Fix:** The archetype K-line is overriding your specs. Change the archetype or add explicit counters.

### Mode-Collapse

Single-agent LLM with weak constraints collapses to bland mean.

**Symptom:** Output is safe, inoffensive, boring.
**Fix:** Use adversarial-committee or add tension/conflict seeds.

---

## Metrics

The Simulator Effect is working when:

1. **Readers imagine more than you wrote**
2. **LLM generates coherent detail you did not specify**
3. **Different interpreters imagine similar worlds**
4. **Sparse specs produce rich output**

---

## See Also

- [k-lines/](../k-lines/) -- The mechanism
- [yaml-jazz/](../yaml-jazz/) -- Comments as seeds
- [adversarial-committee/](../adversarial-committee/) -- Mode-collapse antidote
- [empathic-templates/](../empathic-templates/) -- Smart generation
- [constructionism/](../constructionism/) -- Mental model building
- [sims-astrology.md](../../designs/sims/sims-astrology.md) -- The proof

---

*"The game provides scaffolding. Imagination does the rendering."*

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.