Prompt Injection

BlazeTracker doesn’t just track state — it injects it back into your LLM’s context so the model stays aware of the current scene. This page explains what gets injected, in what format, and how to control it.

“Prompt injection” here refers to injecting tracked state into the LLM prompt, not the security vulnerability. BlazeTracker uses SillyTavern’s official extension prompt API.

What Gets Injected

The injection includes several sections, each wrapped in XML-style tags:

Story So Far

[Story So Far]
Chapter 1: A Chance Meeting
  Two strangers cross paths at the marketplace...
Chapter 2: The Journey Begins
  Having agreed to travel together, they set out...
[/Story So Far]

Recent completed chapters with titles and summaries. Controlled by the Max Recent Chapters setting (default: 5).

Scene State

[Scene State]
Topic: Exploring the ruins
Tone: Cautious, adventurous
Tension: guarded (suspense, escalating)
Time: Thursday, March 15, 1247 at 2:30 PM (day)
Location: Ancient Ruins - Central Chamber - Near the collapsed pillar
Nearby objects: torch, crumbled stone tablet, iron door
Climate: overcast, 48°F, humid
Characters present:
  Alice: crouching near the inscription; doing: translating runes; mood: focused, excited; physical: dusty; wearing: leather vest, cargo pants, hiking boots
  Bob: standing guard at the entrance; doing: watching the corridor; mood: alert, nervous; wearing: chain shirt, dark trousers, boots, short sword on belt
[/Scene State]

The complete current state: time, location, props, climate, and all present characters with their positions, activities, moods, physical states, and outfits.

Recent Events

[Recent Events]
- Alice discovered ancient runes on the chamber wall
- Bob heard footsteps echoing from the lower levels
- Alice identified the runes as a warning about guardians
[/Recent Events]

Narrative events from the current chapter. Controlled by the Max Recent Events setting (default: 15).

Knowledge Gaps

[Knowledge Gaps]
Bob was not present for: Alice discovered a hidden passage behind the altar
Alice was not present for: Bob overheard guards discussing the patrol route
[/Knowledge Gaps]

Events that present characters missed because they weren’t witnesses. This creates natural dramatic irony — the LLM knows what each character doesn’t know, enabling more realistic interactions.

Relationships

[Relationships]
Alice & Bob: friendly
  Alice → Bob: feels trust, curiosity; wants to learn his past
  Bob → Alice: feels protective, respect; hides: his true identity; wants to keep her safe

Alice & Clara: strained
  Alice → Clara: feels betrayal, anger
  Clara → Alice: feels guilt, regret; wants forgiveness
[/Relationships]

Relationship states for all pairs where both characters are present. Includes feelings, wants, and secrets (secrets can be disabled via settings).

Injection Depth

The injection is placed in-chat at a configurable depth. Depth 0 means right before the most recent messages (the default). Higher depth values push the injection further back in the conversation.

This is configured via the Injection Depth setting.

What Controls Each Section

Each section is controlled by its corresponding track toggle:

Section	Track Toggle	Setting
Story So Far	`narrative`	Max Recent Chapters
Scene State (time)	`time`	—
Scene State (location)	`location`	—
Scene State (props)	`props`	—
Scene State (climate)	`climate`	—
Scene State (characters)	`characters`	—
Scene State (topic/tone/tension)	`scene`	—
Recent Events	`narrative`	Max Recent Events
Knowledge Gaps	`characters` + `narrative`	—
Relationships	`relationships`	Include Secrets

If a track toggle is disabled, its corresponding sections are omitted from injection.

Auto-Injection Toggles

You can independently disable auto-injection for state and narrative content:

Setting	Controls	Default
Auto Inject State	Scene State + Relationships	`true`
Auto Inject Narrative	Story So Far + Recent Events + Knowledge Gaps	`true`

These settings are in Advanced Settings > Context Injection.

When you disable auto-injection for a category, BlazeTracker stops inserting that content into prompts automatically. However, extraction still runs normally — the data is still tracked, just not injected.

This is useful when you want to place content manually using ST Macros ({{btState}} and {{btNarrative}}).

Token Budget

The injection content counts against your context window. For a fully-tracked scene with several characters and chapters, the injection might be 500-1500 tokens.

The Token Budget setting lets you cap injection size. When set to 0 (default), it uses SillyTavern’s full context size. Setting a lower value trims older chapters and events first.

ST Macros

BlazeTracker registers two macros that can be placed anywhere in your prompts:

{{btState}} — Current scene state (time, location, characters, relationships)
{{btNarrative}} — Chapter summaries and current chapter events

These work independently of auto-injection — you can use them even when auto-injection is disabled. See the ST Macros guide for details and examples.

Procedural Weather Prefix Caching