NDL Pattern: Combat Narration

Overview

Combat narration in NDL demonstrates the language’s ability to describe complex action sequences with deterministic outcomes. By pre-computing combat results (hits, misses, damage), NDL allows even small 8B-9B parameter models to generate coherent, dramatic combat narrative without hallucinating outcomes.

Core Philosophy

“The game mechanics are the easy part, you just tell the LLM whether it hit or not, and the LLM goes on to make up fancy combat words explaining that you gottem or that you made a fucky wucky” - User-50h100a

Combat in NDL follows the separation of concerns principle:

Game logic determines - Who attacks, what happens, damage dealt
NDL describes - The action, manner, intention, target, and result
LLM narrates - The dramatic prose describing the outcome

Basic Combat Pattern

Simple Attack

do($"player", "attack") ~ "fiercely" -> target($"goblin") -> result("hit") -> damage(8)

Generated Narrative (typical output):

“With fierce determination, you swing your blade at the goblin. Your strike connects solidly, dealing 8 damage. The goblin staggers backward, clutching the wound.”

Attack with Miss

do($"player", "attack") ~ "recklessly" -> target($"guard") -> result("miss")

Generated Narrative:

“You attack recklessly, leaving yourself exposed. The guard easily sidesteps your wild swing, your weapon finding only air.”

Enemy Attack

do($"witch", "attack") ~ "fireball" -> target($"player") -> result("hit") -> damage(5) -> describe("burning")

Generated Narrative:

“The witch hurls a blazing fireball in your direction. The flames strike you squarely, dealing 5 damage. Your skin blisters from the burning heat.”

Advanced Combat Patterns

Multi-Action Combat Sequence

do($"player", "dodge") ~ "quickly" -> result("success")
-> do($"player", "counterattack") ~ "sword" -> target($"bandit") -> result("critical") -> damage(15)

Design Notes:

Each action is explicitly sequenced with ->
Success/failure of defensive actions precedes offensive actions
Critical hits are just another result type that the narrative layer interprets

Weapon-Specific Attacks

do($"player", "swing") ~ "two-handed axe" -> target($"door") -> result("destroyed")

do($"archer", "shoot") ~ "longbow" -> target($"player") -> result("graze") -> damage(2)

Pattern: The manner modifier (~) specifies the weapon or attack method, letting the LLM add weapon-appropriate descriptive language.

Failed Combat Action

From real Gemma 2 testing (see Media/image-11F41.png):

do($"you", "swing makeshift club against table's surface",
   intention="test club's combat readiness")
-> result("failure")
-> system_response("something goes wrong when you try to perform this action")
-> describe()

Generated Narrative:

“You raise the gnarled branch, hefting it awkwardly in your hand. It’s heavier than you anticipated, the rough bark digging into your palm. You bring it back over your head, aiming for the tabletop. But as you swing, your momentum falters. The branch veers off course, catching on the edge of the chair.

With a sickening crack, the makeshift club splinters, breaking into several pieces. The impact sends a shower of splinters flying, scattering across the floor. You stare down at the useless fragments in your hand, your heart sinking.”

Design Notes:

Intention reveals why the player attempted this (testing the weapon)
Result is pre-determined (failure) - no hallucination of success
system_response() adds meta-information about action failure
describe() prompts detailed narrative of the failure’s consequences

Combat Turn Structure

Turn-Based Combat (Basic)

# Player turn
do($"player", "attack") ~ "sword" -> target($"orc") -> result("hit") -> damage(6)
-> wait("enemy response")
 
# Enemy turn
do($"orc", "attack") ~ "club" -> target($"player") -> result("hit") -> damage(4)
-> wait("player response")

Processing Flow:

Player declares action → game resolves → NDL generated → LLM narrates
Wait marker signals turn boundary
Enemy action processed same way

Combat with Positioning

do($"player", "attack") ~ "sword pointing forward"
-> target($"enemy weapon")
-> result("deflected")
-> do($"player", "riposte") ~ "quickly" -> target($"enemy") -> result("hit") -> damage(5)

Pattern Insight (from User-monkeyrithms):

“one thing I experimented with is asking it which way the player’s weapon is pointing as it’s being held (as in fencing, like picture two characters fighting with swords — the swords are usually held in front of you).

it is reminded of that direction when it posts, so if the player’s sword is being held forwards, a prompt reminder will tell the LLM, ‘the player’s weapon is in front of them. This will make a direct attack difficult. You should deal with the weapon, first.’

This substantially raised the coherency of scenes that involved combat with weapons”

Chain-of-Thought Combat Resolution

For complex combat requiring tactical reasoning:

roll("attack", modifiers="+2 flanking") -> result("18")
-> do($"player", "attack") ~ "flanking strike" -> target($"knight") -> result("hit") -> damage(9)

Pattern: Use roll() to expose the game’s decision process before narration, making the outcome feel earned and logical.

Combat Scene Modes

Scene Transition to Combat

do($"goblin", "attack") -> target($"player")
-> wait("combat mode")
-> system_response("combat has begun")

Implementation Note (from User-monkeyrithms):

“when combat has begun. The game window will turn red as a visual indicator, and NPCs will be fed different kinds of prompts that facilitate that sort of scene better.”

“as the combat started the game detected combat and enters a different ‘mode’ where the characters are given different kinds of prompts that have a stronger emphasis on doing only 1 thing per turn, and not describing the outcome of their actions”

Combat-Specific Prompting

During combat mode, the system prompt changes:

Emphasize single actions per turn
Don’t describe outcomes of actions (game determines these)
Focus on character intent and emotional state
Respect weapon positioning and tactical constraints

Combat State Tracking

Status Effects

do($"wizard", "cast") ~ "slow spell" -> target($"warrior") -> result("success")
-> describe("warrior is slowed")

do($"player", "attack") ~ "while slowed" -> target($"enemy") -> result("miss")
-> system_response("slowed condition affects attack accuracy")

Health Thresholds

do($"enemy", "attack") -> target($"player") -> result("hit") -> damage(10)
-> system_response("player health critical")
-> wait("player decision")

Pattern: Use system_response() to inject state information that influences narrative tone (desperation, caution, triumph).

Common Anti-Patterns

❌ Don’t: Let LLM Decide Combat Outcomes

# WRONG - leaves outcome to LLM
do($"player", "attack") -> target($"dragon")
# LLM might hallucinate a hit when game logic determined a miss

✅ Do: Always Specify Result

# CORRECT - outcome is deterministic
do($"player", "attack") -> target($"dragon") -> result("miss")
# LLM narrates the miss creatively, but can't change the outcome

❌ Don’t: Vague Action Descriptions

# WRONG - too generic
do($"player", "fight")

✅ Do: Specific Actions with Context

# CORRECT - specific action with manner
do($"player", "thrust") ~ "spear" -> target($"shield") -> result("blocked")

❌ Don’t: Mix Multiple Targets Ambiguously

# WRONG - unclear which enemy is targeted
do($"player", "attack") -> target($"enemies")

✅ Do: One Action, One Target

# CORRECT - explicit target
do($"player", "attack") -> target($"goblin_leader") -> result("hit") -> damage(7)

Integration with Game Systems

D&D-Style Combat

def generate_combat_ndl(attacker: Entity, defender: Entity, attack_roll: int,
                        ac: int, damage_roll: int) -> str:
    """Generate NDL for D&D-style combat resolution."""
    hit = attack_roll >= ac
 
    # Build NDL based on game logic
    ndl = f'do(${attacker.id}, "attack")'
    ndl += f' ~ "{attacker.equipped_weapon.name}"'
    ndl += f' -> target(${defender.id})'
 
    if hit:
        ndl += f' -> result("hit") -> damage({damage_roll})'
        if attack_roll >= ac + 10:  # Critical hit
            ndl += ' -> describe("critical strike")'
    else:
        ndl += f' -> result("miss")'
        if attack_roll == 1:  # Critical fumble
            ndl += ' -> describe("fumble")'
 
    return ndl

Action Point System

# First action (2 AP)
do($"player", "power attack") ~ "greatsword" -> target($"golem") -> result("hit") -> damage(12)
-> system_response("2 action points spent, 1 remaining")
 
# Second action (1 AP)
-> do($"player", "move") ~ "defensive position"
-> system_response("turn complete")

Damage Types and Resistance

do($"player", "cast") ~ "fire bolt" -> target($"ice elemental") -> result("hit") -> damage(8)
-> describe("weak to fire, extra damage")
-> system_response("ice elemental takes 16 damage total")

Genre-Specific Variations

Fantasy Melee Combat

do($"knight", "charge") ~ "lance" -> target($"orc chieftain")
-> result("hit") -> damage(14) -> describe("mounted charge")

Sci-Fi Firefight

do($"player", "fire") ~ "plasma rifle" -> target($"alien")
-> result("hit") -> damage(6) -> describe("shields at 40%")

Horror Combat (Futility)

do($"player", "shoot") ~ "pistol" -> target($"monster")
-> result("hit") -> damage(2) -> system_response("it barely notices")

Tactical Combat (Positioning Matters)

do($"player", "flank") ~ "stealthily" -> target($"guard") -> result("advantage")
-> do($"player", "backstab") ~ "dagger" -> target($"guard") -> result("critical") -> damage(18)

Performance Considerations

Model Requirements

Minimum: 8B-9B parameter models (Gemma 2, Llama 3)
Optimal: 13B+ for complex combat with multiple actors
Works well on: Local inference, no GPT-4 required

Token Efficiency

Combat NDL is token-efficient:

Simple attack: ~15-25 tokens of NDL
Complex sequence: ~40-60 tokens
Generated output: 50-150 tokens typically

Compare to pure prompt-based combat:

Prompt with full context: 200-400 tokens
Risk of hallucination: High without NDL structure

Testing Combat Patterns

Validation Checklist

Hit results generate successful attack narrative
Miss results never show successful hits
Damage values appear in narrative accurately
Weapon types influence descriptive language
Status effects mentioned in follow-up narration
Critical hits generate appropriately dramatic language
Multiple combatants tracked correctly across turns
Combat → non-combat transitions handled smoothly

Common Edge Cases

Overkill damage: Player deals 50 damage to enemy with 5 HP

do($"player", "smash") ~ "hammer" -> target($"rat") -> result("overkill") -> damage(50)
-> describe("obliterated")

Healing during combat:

do($"cleric", "cast") ~ "healing word" -> target($"warrior") -> result("healed") -> damage(-8)
# Negative damage = healing

Environmental damage:

system_response("lava floor deals damage") -> damage(3) -> target($"player")

Real-World Implementation

From User-monkeyrithms ReallmCraft implementation:

Combat detected automatically when attack actions occur
Different prompt templates activate for combat vs exploration
Turn-based structure enforced via wait() sequencing
Character death triggers scene transition to “Afterlife” location

From User-veritasr DirectorAPI:

Combat NDL parsed via regex detection of do(, target(, result(
Each action in sequence triggers separate LLM generation pass
Results aggregated before player sees final narrative

sequencing - Chaining combat actions with ->
result - Deterministic outcome specification
manner-modifier - Weapon and style descriptions
intention - Why character is attacking (intimidate, defend, etc.)
scene-transitions - Moving into/out of combat mode
wait - Turn boundaries and response timing

LLM World Engine Knowledge Base

Explorer

combat-narration

NDL Pattern: Combat Narration

Overview

Core Philosophy

Basic Combat Pattern

Simple Attack

Attack with Miss

Enemy Attack

Advanced Combat Patterns

Multi-Action Combat Sequence

Weapon-Specific Attacks

Failed Combat Action

Combat Turn Structure

Turn-Based Combat (Basic)

Combat with Positioning

Chain-of-Thought Combat Resolution

Combat Scene Modes

Scene Transition to Combat

Combat-Specific Prompting

Combat State Tracking

Status Effects

Health Thresholds

Common Anti-Patterns

❌ Don’t: Let LLM Decide Combat Outcomes

✅ Do: Always Specify Result

❌ Don’t: Vague Action Descriptions

✅ Do: Specific Actions with Context

❌ Don’t: Mix Multiple Targets Ambiguously

✅ Do: One Action, One Target

Integration with Game Systems

D&D-Style Combat

Action Point System

Damage Types and Resistance

Genre-Specific Variations

Fantasy Melee Combat

Sci-Fi Firefight

Horror Combat (Futility)

Tactical Combat (Positioning Matters)

Performance Considerations

Model Requirements

Token Efficiency

Testing Combat Patterns

Validation Checklist

Common Edge Cases

Real-World Implementation

Related Patterns

Further Reading

Graph View

Table of Contents

Backlinks