Custom Prompts

Custom Prompts

Every extraction prompt in BlazeTracker can be customized. This is useful for tuning extraction accuracy for your specific model, adding instructions for edge cases, or changing the output format.

Prompt Structure

Each extraction prompt has two parts:

System Prompt

Static instructions that define the extractor’s role and output format. This part doesn’t change between extraction calls and is designed to be cached by inference backends (see Prefix Caching).

Example:

You are a state tracker for a roleplay scene. Extract the current
narrative time from the messages below. Output JSON with fields:
year, month, day, hour, minute, second.

User Template

Dynamic content with placeholders that get filled in at extraction time. This changes per call because it includes the actual chat messages and current state.

Example:

Messages:
{{messages}}

Previous time: {{previousTime}}

What is the current narrative time?

Available Placeholders

Each prompt has its own set of available placeholders. Common ones include:

PlaceholderDescription
{{messages}}Recent chat messages
{{characterName}}AI character’s name
{{userName}}User’s persona name
{{previousState}}Previous extracted state for context
{{characters}}List of present characters
{{worldinfo}}Lorebook/world info entries (if enabled)

The exact placeholders available for each prompt are shown in the settings UI when you expand that prompt’s customization section.

Temperature Overrides

Each extractor category has a default temperature:

CategoryDefaultRationale
Time0.3Low for deterministic time parsing
Location0.5Moderate for reasonable location inference
Props0.5Moderate
Climate0.3Low for consistent classification
Characters0.5Moderate
Relationships0.6Slightly higher for nuanced feelings
Scene0.5Moderate
Narrative0.6Slightly higher for creative summaries

You can override temperatures at two levels:

  1. Category temperature — Changes the default for all prompts in that category (Settings > Advanced > Category Temperatures)
  2. Per-prompt temperature — Overrides the category default for a specific prompt

Tips for Custom Prompts

For smaller models (27B-32B)

  • Be more explicit about output format
  • Add example outputs
  • Reduce the number of fields requested
  • Lower temperatures for more deterministic output

For larger models (70B+)

  • Default prompts usually work well
  • Can increase temperatures slightly for more nuanced extraction
  • Can add additional instructions for edge cases

Common customizations

  • Time extraction: Add timezone handling if your story involves travel
  • Character extraction: Add species-specific outfit rules (e.g., “tail slot” for non-human characters)
  • Relationship extraction: Adjust sensitivity for relationship status changes
  • Narrative extraction: Change the style of event summaries

Prompt Prefix and Suffix

The Prompt Prefix and Prompt Suffix settings (under Advanced) are prepended/appended to the user template of all prompts. Common use cases:

  • /nothink prefix for models that support disabling chain-of-thought
  • Additional system-level instructions that apply to all extractors