Template Pattern Cheat Sheet

Scene Context Inclusion

Include scene history with budget control:

{% set history = scene.context_history(budget=budget, min_dialogue=20, sections=False) %}
<|SECTION:SCENE|>
{% for scene_context in history -%}
{{ scene_context }}
{% endfor %}
<|CLOSE_SECTION|>

Variables explained:

• budget - Maximum token allocation for context (max_tokens will be available and hold the total token budget)
• min_dialogue - Minimum number of dialogue lines to include
• scene - The current scene object (automatically available)

Reusable template: Include scene-context.jinja2 instead of writing this manually.

Memory Context Inclusion

Retrieve relevant memories based on context:

{% set memory_stack = agent_action("narrator", "rag_build", prompt=memory_prompt, sub_instruction=memory_goal) %}
{% if memory_stack %}
<|SECTION:POTENTIALLY RELEVANT INFORMATION|>
{%- for memory in memory_stack -%}
{{ memory|condensed }}
{% endfor -%}
<|CLOSE_SECTION|>
{% endif %}

Variables explained:

• memory_prompt - The text to search memories for
• memory_goal - Optional specific goal for memory retrieval
• agent_action() - Built-in function to call agent actions

Reusable template: Include memory-context.jinja2 for standard memory inclusion.

Character Context Inclusion

Include character information with conditional detail:

<|SECTION:CHARACTERS|>
{% for character in scene.characters %}
### {{ character.name }}
{% if max_tokens > 6000 -%}
{# we have lots of space so print the whole character sheet #}
{{ character.sheet }}
{% else -%}
{# we have limited space so print the bare minimum and then query memory for personality #}
{# bare minimum attributes that are guaranteed to be available #}
{{ character.filtered_sheet(['name', 'gender', 'age']) }}
{# query memory for personality #}
{{ query_memory("what is "+character.name+"'s personality?", as_question_answer=False) }}
{% endif %}
{{ character.description }}
{% endfor %}
<|CLOSE_SECTION|>

Variables explained:

• scene.characters - List of all characters in the scene
• max_tokens - Available token budget (from agent client)
• character.sheet - Full character attributes
• character.filtered_sheet() - Subset of character attributes
• query_memory() - Built-in function to query memory

Reusable template: Include character-context.jinja2 for standard character listing.

Including other templates

{% include "memory-context.jinja2" %}
{% with budget=budget %}{% include "scene-context.jinja2" %}{% endwith %}

The {% include %} directive allows you to reuse existing templates. Common includes:

• scene-context.jinja2 - Standard scene history
• memory-context.jinja2 - Memory retrieval
• character-context.jinja2 - Character listings
• extra-context.jinja2 - Additional context like pins and reinforcements
• scene-intent.jinja2 - Scene intention

Conditional sections based on token budget

{% if max_tokens > 6000 -%}
  {{ detailed_content }}
{% else -%}
  {{ condensed_content }}
{% endif %}

Response scaffolding

You can use the {{ bot_token }} variable to separate the prompt and any response scaffolding you want to do.

Scaffolding?

Scaffolding is a technique where you provide the beginning of the response, and then let the agent continue.

This can be useful when you want to coerce the agent down a specific path or formatting.

<|SECTION:TASK|>
Analyze the current moment in the scene and then make a prediction about what will happen next.

Do your analysis first!

Provide your response in the following format:

ANALYSIS: .. your analysis ..

PREDICTION: .. your prediction ..
<|CLOSE_SECTION|>
{{ bot_token }}ANALYSIS:

Its important to note that with this approach, anything you put after the {{ bot_token }} will be missing from the response. So in this example the ANALYSIS: string will not be part of the response.

If you need it, you can instead use the set_prepared_response function.

<|SECTION:TASK|>
Analyze the current moment in the scene and then make a prediction about what will happen next.

Do your analysis first!

Provide your response in the following format:

ANALYSIS: .. your analysis ..

PREDICTION: .. your prediction ..
<|CLOSE_SECTION|>
{{ set_prepared_response("ANALYSIS: ") }}

Response scaffolding for data structures

Scaffolding can be exceptionally helpful if you are prompting the agent to generate a data structure and nothing else, because you can simply coerce it to start with the data structure you want.

<|SECTION:TASK|>
Generate a list of objects in the scene, provide a name and description for each object.


  ```json
  {
    "objects": [
      {
        "name": "an old shelf",
        "description": "a shelf that is old and has a few scratches on it"
      },
      {
        "name": "a dirty carpet",
        "description": "a carpet that is dirty and has a few stains on it"
      }
    ]
  }
  ```

<|CLOSE_SECTION|>
{{ set_data_response({"objects": [{"name":""}]}) }}