Documentation for how system prompts are dynamically assembled, structured, and sent to LLMs.
The system prompt is the foundational instruction set that defines the AI's persona, capabilities, tools, and operational constraints. In SafeAppeals, this prompt is dynamically generated at runtime based on:
- Chat Mode (
case_manager
, research
, drafting
, etc.)
- Available Tools (Built-in + MCP)
- Workspace Context (Open files, directory structure)
- User Settings (Custom instructions, .fileorg.json)
The construction of the prompt is orchestrated by the browser process before the message is sent to the main process.
graph TD
subgraph "Browser Process (Services)"
Service[ConvertToLLMMessageService<br>src/.../browser/convertToLLMMessageService.ts]
end
subgraph "Common (Prompt Logic)"
Entry[chat_systemMessage<br>src/.../common/prompt/prompts.ts]
subgraph "Core Instructions"
SysPrompt[getSystemPrompt<br>src/.../common/prompt/systemPrompt.ts]
end
subgraph "Tool Definitions"
ToolGen[systemToolsXMLPrompt<br>src/.../common/prompt/prompts.ts]
ToolsObj[builtinTools Object<br>src/.../common/prompt/prompts.ts]
Schemas[toolSchemas.ts<br>src/.../common/prompt/toolSchemas.ts]
end
end
%% Execution Flow
Service -->|1. Calls with Context| Entry
Entry -->|2. Request Text| SysPrompt
Entry -->|3. Request XML| ToolGen
%% Internal Dependencies
ToolGen -->|4. Read Definitions| ToolsObj
ToolsObj -.->|5. Import Description| Schemas
%% Return Flow
SysPrompt -->|Return| Entry
ToolGen -->|Return| Entry
Entry -->|Return Full String| Service
Location:
src/vs/workbench/contrib/void/browser/convertToLLMMessageService.ts
This service is the entry point. It gathers live data from the IDE:
- List of open files (
openedURIs
)
- Active editor (
activeURI
)
- Workspace folders
- Persistent terminal IDs
- Directory structure string
It then calls
chat_systemMessage
with this context.
Location:
src/vs/workbench/contrib/void/common/prompt/prompts.ts
This file contains the logic to stitch everything together.
chat_systemMessage
: The main function that combines the base prompt with tool definitions.systemToolsXMLPrompt
: Generates the XML block for available tools and guidelines.builtinTools
: Defines the name, description, and parameters for every tool.
Location:
src/vs/workbench/contrib/void/common/prompt/systemPrompt.ts
This file contains the static text and logic for the AI's behavior. It exports
getSystemPrompt
, which returns the prompt string based on the
ChatMode
.
Sections include:
- Identity & Purpose: Role definition.
- Response Style: Communication guidelines.
- Mode Workflows: Specific instructions for
case_manager
, research
, etc.
- Tool Calling Format: ANTML instructions (
<function_calls>
).
- Policy Verification: Rules for citing sources.
Location:
src/vs/workbench/contrib/void/common/prompt/toolSchemas.ts
Contains complex JSON schemas for tools like
edit_document
(DOCX/XLSX editing operations). These schemas are imported by
prompts.ts
to populate the tool descriptions.
The final string sent to the LLM follows this order:
- Identity & Purpose ("You are an expert...")
- Response Style ("Be direct, no self-narration...")
- Mode-Specific Workflow (e.g., "Research Mode: RAG-First...")
- Tool Calling Format (ANTML instructions & examples)
- Available Tools (List of tools in XML format)
- Context Management (Token usage guidelines)
- System Environment (OS, Workspace folders, Open files)
- Feature-Specific Instructions (Timeline, Case Config, Documents)