Prompt System

The prompt system manages system prompt composition through profiles, sections, and a global prompt. All components combine to produce the final system prompt sent to the model.

Composition Pipeline

graph LR
    GP[Global Prompt] --> C[Composer]
    MCP[MCP Tools Prompt<br/>auto-generated] --> C
    P[Profile Sections<br/>ordered, toggleable] --> C
    AC[Additional Context<br/>memory, overrides] --> C
    C --> SP[Final System Prompt]

Composition order:

1. Global prompt — Core persona and behavior (applied to all interactions)
2. MCP tools prompt — Auto-generated description of available tools
3. Profile sections — Enabled sections sorted by order
4. Additional context — Memory context, request-specific overrides

If system_override is set in the request, it replaces the entire composed prompt.

Profiles

A profile is a named collection of sections representing a use case (e.g., “Developer”, “Creative Writer”).

# system_prompts.yaml
profiles:
  - id: default
    name: Default Assistant
    description: General-purpose AI assistant
    is_default: true
    sections:
      - id: identity
        name: Identity
        type: persona
        content: "You are a helpful AI assistant."
        enabled: true
        order: 0
      - id: guidelines
        name: Guidelines
        type: constraints
        content: "Be concise. Cite sources when possible."
        enabled: true
        order: 1

Profiles are selected via the profile_id parameter in chat endpoints. The default profile is used when none is specified.

Section Types

Sections can be individually enabled/disabled and reordered within a profile.

Global Prompt

The global prompt applies to all interactions regardless of profile. It defines the core persona that persists across profile switches.

manager = get_prompt_manager()
manager.set_global_prompt("Always be helpful and concise.", enabled=True)

MCP Tools Prompt

When MCP servers are connected, the system auto-generates a tools description that gets injected into the system prompt. This tells the model what tools are available and how to use them.

Generated by generate_mcp_tools_prompt() from the list of connected tools.

Storage

Prompts are stored in data/system_prompts.yaml. The PromptManager singleton:

• Initializes with defaults from prompts/defaults.py
• Loads custom configuration from YAML on startup
• Saves changes back to YAML on profile/global prompt updates

API Endpoints

See API Endpoints for full details.

Related

• API Models: Prompt — PromptProfile, PromptSection, GlobalPrompt schemas
• Chat — How prompts integrate with chat requests
• Config file: data/system_prompts.yaml