LLM World Engine Knowledge Base

00-PROMPT-INDEX

7 min read

Prompt Library Index

Overview

This library catalogs all prompts, prompt techniques, and prompting strategies discussed in the LLM World Engine Discord community. The collection represents nearly two years of experimentation, refinement, and practical implementation across multiple LLM game engine projects (ReallmCraft, ChatBot RPG).

Total Prompts Documented: 17 complete Techniques Covered: 9 Models Tested: GPT-4, GPT-3.5, Claude, Mixtral, Llama 3, Gemma 2, Mistral, EstopianMaid, Stheno Date Range: January 2024 - December 2025

Core Philosophy

The community’s prompt engineering approach centers on a fundamental insight:

LLMs excel at narration, not decision-making.

Rather than asking LLMs to track state and make game logic decisions (which they do poorly), prompts should provide structured descriptions of predetermined events and ask the LLM to translate them into natural narrative prose.

Quick Reference by Category

Narration Prompts

Transform programmatic game events into natural language narrative:

Generation Prompts

Create game content (characters, locations, items, quests):

Constraint Prompts

Control and limit LLM behavior to prevent hallucinations:

Retrieval Prompts

RAG and memory system optimization:

Reasoning Prompts

Multi-step thinking and decision support:

System Prompts

Base instructions that define LLM behavior:

Quick Reference by Technique

1. NDL (Natural Description Language)

Status: Production-ready, PRIMARY TECHNIQUE Creator: User-veritasr File: narration/ndl-to-narrative.md

Convert programmatic game events into structured markup that LLMs reliably translate into narrative prose. Works on 7B-9B models.

2. Chain of Thought (CoT)

Status: Proven File: reasoning/chain-of-thought.md

Force LLM to think step-by-step before responding. Especially useful for smaller models.

3. Few-Shot Prompting

Status: Essential for new formats File: techniques/few-shot-examples.md

Provide 1-3 examples to teach LLMs custom output formats like NDL.

4. Constraint-Based Prompting

Status: Critical for reliability File: constraint/anti-hallucination.md

Tell LLMs what they CANNOT do to prevent hallucinations.

5. Template-Based Generation

Status: Production-ready Files: generation/character-generation.md, generation/location-generation.md

Use structured templates with placeholders for consistent content generation.

6. Context Inheritance

Status: Proven pattern File: generation/location-generation.md

Child elements (locations, NPCs) inherit thematic properties from parent regions.

7. HyDE (Hypothetical Document Embeddings)

Status: Proven for RAG File: retrieval/query-formulation-hyde.md

Generate hypothetical questions content could answer, improving semantic search retrieval quality by 40-80%.

8. Meta-Prompting

Status: Production-ready File: generation/template-generation.md

Prompts that generate other prompts. Use LLMs to create structured templates for procedural content generation.

9. Binary Classification

Status: Production-ready File: reasoning/binary-classification.md

Extract binary decisions from natural language for game logic. Question tree approach with [YES]/[NO] parsing enables deterministic state management.

Files Included in This Library

Narration (5 files)

  • narration/ndl-to-narrative.md - PRIMARY: NDL markup → natural prose
  • narration/scene-description.md - Location and environmental descriptions
  • narration/action-narration.md - Player/NPC action narration
  • narration/dialogue-generation.md - Character speech with subtext
  • narration/combat-narration.md - Turn-based combat sequences

Generation (3 files)

  • generation/character-generation.md - Create NPCs with full character sheets
  • generation/location-generation.md - Generate locations with context inheritance
  • generation/template-generation.md - NEW: Meta-prompts that generate templates

Constraints (4 files)

  • constraint/anti-hallucination.md - Core constraint rules approach
  • constraint/hallucination-prevention.md - Explicit prohibitions with data grounding
  • constraint/format-enforcement.md - NEW: Structured output control with brackets
  • constraint/length-limiting.md - NEW: Token budget management (170-token technique)

Note: Both anti-hallucination.md and hallucination-prevention.md address the same core problem (preventing LLM hallucinations) but use complementary approaches. The former focuses on rule-based constraints, while the latter emphasizes explicit data grounding and comprehensive prohibitions. Use both together for maximum effectiveness.

Retrieval (1 file)

  • retrieval/query-formulation-hyde.md - NEW: HyDE for improved semantic search

Reasoning (2 files)

  • reasoning/chain-of-thought.md - Step-by-step reasoning patterns
  • reasoning/binary-classification.md - NEW: Yes/No validation for game state extraction

System (1 file)

  • system/narration-engine-system.md - Base narration engine instructions

Techniques (1 file)

  • techniques/few-shot-examples.md - Teaching by example patterns

Best Practices Summary

DO:

  • ✅ Use NDL for narration (most reliable technique)
  • ✅ Separate decision-making from narration
  • ✅ Test on target model early and often
  • ✅ Apply constraints liberally
  • ✅ Use few-shot for new formats
  • ✅ Validate and post-process outputs
  • ✅ Cache generated content

DON’T:

  • ❌ Let LLM make game logic decisions
  • ❌ Assume LLM remembers implicit rules
  • ❌ Over-complicate prompts
  • ❌ Give unlimited creative freedom
  • ❌ Use GPT-4 prompts on 7B models unchanged

Model Recommendations

Small Models (7B-9B)

Best for: Narration with NDL Requirements: Tight constraints, few-shot examples, simple prompts Temperature: 0.6-0.9 for narration

Medium Models (13B-30B)

Best for: Balanced narration and generation Requirements: Moderate constraints, some examples Temperature: 0.5-0.8

Large Models (GPT-4, Claude)

Best for: Complex generation, reasoning Requirements: Can work with just instructions Temperature: 0.3-0.7 depending on task

Temperature Guide

Structured Output:  0.1 - 0.3  (parsing, validation)
Narration:          0.6 - 0.9  (events, descriptions)
Dialogue:           0.7 - 1.0  (character speech)
Generation:         0.7 - 0.9  (new content)
Reasoning:          0.3 - 0.5  (analysis, planning)

Key Community Insights

“Turns out that when you take away decision making from the LLM it behaves much better.” - User-veritasr

“LLMs are super cliche and shallow on their own devices… but so would we if we just one-shot everything” - User-appl2613

“{{char}} is a logical and realistic text adventure game. Impossible actions must fail.” - User-yukidaore

“gpt-4 was very good, its the only model that can kinda-sorta one-shot a good RP with all the rules just added to the context” - User-monkeyrithms

Evolution Timeline

January 2024: Experimentation with LLM as state manager (failed), CoT patterns February-April 2024: Constraint-based approaches, template prompts, early NDL May-June 2024: NDL formalization, quest pacing, turn counting July 2024-2025: Advanced techniques, character psychology, cross-model testing

Contributing

When adding new prompts:

  1. Use the standard template format (see existing files)
  2. Include effectiveness notes from real usage
  3. Specify which models were tested
  4. Provide complete, working examples
  5. Tag with relevant techniques
  6. Link to source discussions

Usage Notes

All prompts are production-tested in real game engines (ReallmCraft, ChatBot RPG). They represent nearly two years of community experimentation and refinement.

Primary Technique: NDL (Natural Description Language) is the community’s consensus approach for reliable narration. Start here.

For Small Models: Use NDL + constraints + few-shot examples. This combination works reliably on 7B-9B parameter models.

For Content Generation: Use template-based approaches with validation and post-processing.

Last Updated: 2026-01-17 (corrected file count and added hallucination-prevention) Source: LLM World Engine Discord (Jan 2024 - Dec 2025) Total Messages Analyzed: 12,109

Recent Additions (2026-01-17)

Latest Update (2026-01-17 Evening)

  1. Binary Classification (reasoning/binary-classification.md)
    • Yes/No validation prompts for extracting game state from natural language
    • Question tree pattern for multi-step validation
    • Grammar-constrained output forcing
    • Production-tested in ChatBot RPG with 95%+ accuracy
    • Complete implementation examples with Python code

Earlier Additions (2026-01-17 Morning)

  1. Template Generation (generation/template-generation.md)

    • Meta-prompts that generate structured JSON templates
    • Create generators for locations, items, characters, quests
    • Supports hierarchical JITG (Just-In-Time Generation)
    • Production-tested in ReallmCraft

  2. Length Limiting (constraint/length-limiting.md)

    • 170-token sweet spot technique from ChatBot RPG
    • Combines API limits + prompt instructions + regex post-processing
    • Prevents rambling and maintains consistent pacing
    • Based on front-loaded coherence principle

  3. Format Enforcement (constraint/format-enforcement.md)

    • Bracketed output for reliable parsing: [ITEM1],[ITEM2]
    • Structured extraction of items, locations, actions
    • Production-proven in ChatBot RPG inventory system
    • Includes hallucination mitigation strategies

  4. Query Formulation (HyDE) (retrieval/query-formulation-hyde.md)

    • Hypothetical questions for improved semantic search
    • 40-80% better recall vs. standard RAG
    • Hybrid retrieval combining exact, tag, semantic, and graph methods
    • Includes Reciprocal Rank Fusion (RRF) for result combination

Graph View

Backlinks