chain-of-thought

Prompt: Chain of Thought (CoT)

Metadata

• Category: Reasoning
• Technique: Chain of Thought
• Model Tested: All models (especially beneficial for smaller models)
• Contributor: User-50h100a, community adoption
• Status: Proven technique

Purpose

Force the LLM to “think through” a response step-by-step before generating the final output. This reduces impulsive/random responses, improves consistency, and helps the model consider context more carefully.

Particularly useful for:

• State-aware narration
• Validation tasks
• Multi-step reasoning
• Reducing hallucinations

Core Concept

“You just tell the llm to start every response with x,y, and z. It’ll ‘think’ for itself.” - User-underscore_x

The LLM naturally wants to “fill in the blank.” By structuring prompts with incomplete patterns, you force the model to complete them before continuing, creating an implicit reasoning chain.

Template

Basic CoT Pattern

Task: {{TASK_DESCRIPTION}}

Before responding, consider:
1. {{CONSIDERATION_1}}
2. {{CONSIDERATION_2}}
3. {{CONSIDERATION_3}}

Now, {{FINAL_INSTRUCTION}}

Implicit CoT (Fill in the Blank)

Character: {{CHARACTER_NAME}}
Location: {{LOCATION}}
Mood: {{MOOD}}
What {{CHARACTER_NAME}} wants: {{GOAL}}
What {{CHARACTER_NAME}} will say:

The prompt ends here, forcing the LLM to complete the pattern.

Variables

Variable	Description	Example
{{TASK_DESCRIPTION}}	What you need the LLM to do	”Narrate the character’s action”
{{CONSIDERATION_N}}	Step in the reasoning chain	”Check if character has the ability”
{{CHARACTER_NAME}}	Name of character	”Bob”
{{LOCATION}}	Current location	”Tavern”
{{MOOD}}	Character’s emotional state	”Nervous”
{{GOAL}}	Character’s current intention	”Information about the quest”
{{FINAL_INSTRUCTION}}	What to output	”write the dialogue”

Usage Example

Example 1: Statblock-Based Narration

Prompt:

Character: Theron the Warrior
Location: Dark Forest
Current Health: 45/100
Stamina: Exhausted
Mood: Anxious
Recent Events: Just escaped from bandits, wounded
Equipment: Sword (damaged), leather armor
Goal: Find shelter before nightfall

What Theron will do next:

Output:

Theron scans the treeline desperately, looking for any sign of refuge. His wounded leg throbs with each step, but fear of pursuit keeps him moving. He needs shelter, and fast.

Example 2: Action Validation with CoT

Prompt:

Player attempts: "I cast a fireball spell"

Character Sheet:
- Class: Warrior
- Abilities: Sword proficiency, shield bash
- Magic abilities: None

Before responding, check:
1. Does the character have this ability? [No, warrior with no magic]
2. Is this action possible given their abilities? [No]
3. What is the appropriate outcome? [Action fails, explain why]

Narrate the outcome:

Output:

You raise your hands and shout dramatic words, but nothing happens. You're a warrior, not a mage - you have no magical training or innate spellcasting ability.

Example 3: NPC Dialogue with CoT

Prompt:

NPC: Marcus the Innkeeper
Personality: Cautious, protective of regular customers
Current situation: Stranger (player) asking about local bandits
What Marcus knows: The bandits are dangerous, his daughter was threatened
What Marcus thinks: This stranger might be trouble, or might help
Marcus's goal: Assess if player can be trusted

What Marcus is thinking:
[CoT space for model to reason]

What Marcus says:

Output:

What Marcus is thinking:
This one's got the look of a fighter... Could be another bandit, or could be someone who can help. Best to be careful. I'll hint at the problem without saying too much.

What Marcus says:
"Bandits? Aye, they're around. Most folk here learned to keep their heads down and their doors locked." He studies you carefully, wiping the same spot on the bar twice.

Effectiveness Notes

What Works Well

• ✅ Reduces impulsive/random responses
• ✅ Forces consideration of context
• ✅ Especially helpful for smaller models (7B-9B)
• ✅ Creates implicit “reasoning” without explicit instruction
• ✅ Improves consistency across turns
• ✅ Natural way to structure prompts

Known Limitations

• ⚠️ Slower generation (more tokens to produce)
• ⚠️ Increases token usage
• ⚠️ Can be verbose if not constrained
• ⚠️ Requires careful prompt structure
• ⚠️ Some models might skip the reasoning step

Community Feedback

50h100a: “CoT with external state tracking, essentially”

monkeyrithms: “One-shotting combat doesn’t work too well unless it’s GPT-4 in my experience, but a CoT type thing does”

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

CoT Patterns

1. Explicit Reasoning Steps

Before narrating, answer:
1. What is the current situation?
2. What is the character trying to do?
3. Is this action possible?
4. What is the most likely outcome?

Now narrate:

2. State-Checking Pattern

Current State:
- Health: {{HEALTH}}
- Location: {{LOCATION}}
- Equipment: {{EQUIPMENT}}
- Recent events: {{EVENTS}}

Given this state, describe what happens next:

3. Three-Question Framework (TTRPG-Inspired)

Action: {{PLAYER_INPUT}}

What (are they doing):
How (are they doing it):
Why (are they doing it):

Based on these, narrate the outcome:

This is the foundation of NDL’s design philosophy.

4. Validation Chain

Player action: {{ACTION}}

Validation:
- Does character have required ability? {{CHECK_ABILITY}}
- Does character have required item? {{CHECK_ITEM}}
- Is action physically possible? {{CHECK_PHYSICS}}
- Does action follow game rules? {{CHECK_RULES}}

Result:

5. NPC Response Chain

NPC: {{NPC_NAME}}
Player said: {{PLAYER_DIALOGUE}}

NPC's internal reaction:
[What they actually think]

NPC's external response:
[What they choose to say]

Creates natural subtext in dialogue.

Integration with Other Techniques

CoT + NDL

NDL Events:
do($"player", "persuade")~"honeyed words"
->target($"guard")
->roll("charisma", result="failure")

Before narrating, consider:
- The player attempted persuasion
- They used a charming approach
- The dice roll failed
- How would the guard realistically react?

Narrate the outcome:

CoT + Constraints

Player action: "I jump to the rooftop"

Constraint check:
- Character sheet: Normal human, no special abilities
- Physics: Rooftop is 15 feet high
- Conclusion: Impossible without equipment

Narrate the failed attempt:

CoT + Few-Shot

Example 1:
Action: "I search the desk"
What I know: Desk has a hidden compartment
Perception check: Success
Outcome: [Narrate finding the compartment]

Example 2:
Action: "I search the desk"
What I know: Desk has a hidden compartment
Perception check: Failure
Outcome: [Narrate not finding anything unusual]

Now apply this pattern:
Action: {{PLAYER_ACTION}}
What I know: {{WORLD_STATE}}
Relevant check: {{CHECK_RESULT}}
Outcome:

Temperature Settings

For CoT prompts:

• Reasoning portion: 0.3 - 0.5 (want logical, consistent thinking)
• Narration portion: 0.6 - 0.9 (want creative prose)

Some models allow setting different temperatures for different parts of generation.

Advanced: Hidden CoT

You can use CoT internally without showing it to the player:

# Generate with CoT
full_prompt = f"""
Character state: {state}
Action: {action}
 
Internal reasoning:
[Let model think here]
 
Player-facing narration:
"""
 
response = llm.generate(full_prompt)
 
# Parse and only show narration part
narration = extract_section(response, "Player-facing narration:")
return narration

This gives you CoT benefits without verbose output.

When to Use CoT

Good Use Cases:

• ✅ Complex decisions with multiple factors
• ✅ State validation
• ✅ NPC dialogue (thinking vs. saying)
• ✅ Action resolution with multiple checks
• ✅ Smaller models that need guidance
• ✅ Situations requiring consistency

Poor Use Cases:

• ❌ Simple, straightforward narration
• ❌ When token budget is extremely tight
• ❌ Real-time responses (CoT is slower)
• ❌ When you’re using NDL (already structured)

Debugging with CoT

CoT makes debugging easier because you can see the model’s “reasoning”:

Action: {{PLAYER_ACTION}}

Debug trace:
1. Parsed action as: [parse]
2. Checked against abilities: [check]
3. Determined outcome: [outcome]
4. Reason: [explanation]

Narration:

This helps identify where things go wrong.

Related Prompts

• statblock-cot - Specific pattern for state-aware narration
• ndl-to-narrative - Alternative that structures input instead of thinking
• query-formulation - CoT for RAG queries

Source

• Discussion context: Early 2024 architecture discussions
• Key contributor: 50h100a’s statblock experiments
• Referenced in: 02-Prompt-Engineering
• Also called: “CoT with external state tracking”