Prompt Modes

+15 Mana ✨
OpenClaw Prompt Modes Reference(DOCUMENTATION)OpenClaw Sub-Agent Guide(DOCUMENTATION)

Introduction

Not every agent needs the full system prompt. OpenClaw provides three prompt modes that control how much of the system prompt is loaded: Full, Minimal, and None. Choosing the right mode for each agent optimizes performance, reduces token usage, and ensures sub-agents do not inherit unnecessary context from their parent.

Key Concepts

  • Full Mode: Loads the complete system prompt including all bootstrap files, tools, skills, and documentation. Used for primary agents.
  • Minimal Mode: Loads only essential identity and safety sections. Strips out tools, skills, documentation, and workspace context. Used for sub-agents.
  • None Mode: Loads only the identity section (IDENTITY.md). No safety overrides, no tools, no memory. Used for highly specialized sub-agents with custom instructions.
  • Sub-Agent: An agent spawned by another agent to handle a subtask. Sub-agents inherit less context to stay focused and token-efficient.
  • Mode Selection: Configured per agent in the agent definition, not globally.

Real World Context

A primary agent receives a complex request: refactor a module, update tests, and write documentation. It spawns three sub-agents, one for each task. The refactoring sub-agent needs tools and project context (Minimal mode with selected tools). The test-writing sub-agent needs test frameworks and conventions (Minimal mode). The documentation sub-agent only needs to write prose (None mode with custom instructions). Each sub-agent gets only the context it needs, saving thousands of tokens per interaction.

Deep Dive

Prompt modes are set in the agent configuration. Here is how each mode is configured:

json
{
  "agents": {
    "list": [
      {
        "id": "primary-coder",
        "promptMode": "full",
        "model": "anthropic:claude-sonnet-4-20250514"
      },
      {
        "id": "test-writer",
        "promptMode": "minimal",
        "model": "anthropic:claude-sonnet-4-20250514"
      },
      {
        "id": "doc-writer",
        "promptMode": "none",
        "model": "openai:gpt-4o-mini"
      }
    ]
  }
}

This configuration defines three agents with different prompt modes. The primary coder gets the full system prompt. The test writer gets a minimal prompt. The doc writer gets none of the standard prompt sections and relies entirely on its own custom instructions.

Here is what each mode includes and excludes:

markdown
| Section              | Full | Minimal | None |
|----------------------|------|---------|------|
| Safety & Compliance  |  Yes |   Yes   |  No  |
| IDENTITY.md          |  Yes |   Yes   | Yes  |
| SOUL.md              |  Yes |    No   |  No  |
| USER.md              |  Yes |    No   |  No  |
| AGENTS.md            |  Yes |    No   |  No  |
| Tool Definitions     |  Yes | Select  |  No  |
| TOOLS.md             |  Yes |    No   |  No  |
| Skills               |  Yes |    No   |  No  |
| Workspace Context    |  Yes |    No   |  No  |
| Documentation        |  Yes |    No   |  No  |
| MEMORY.md            |  Yes |    No   |  No  |
| Runtime State        |  Yes |   Yes   |  No  |

In Minimal mode, the agent still gets safety guidelines and its identity. It can also receive selected tool definitions if explicitly configured. This is useful for sub-agents that need specific tools but not the full workspace context.

None mode is the most stripped-down option. The agent only gets its IDENTITY.md content. Everything else — personality, instructions, tools, memory — is excluded. This mode is for sub-agents where the parent agent provides all necessary context in the task prompt itself.

The token savings between modes can be dramatic:

markdown
Typical token usage by mode:
- Full mode:    80,000 - 150,000 tokens
- Minimal mode: 5,000 - 15,000 tokens
- None mode:    500 - 2,000 tokens

A Full mode agent might use 100K tokens just for the system prompt, leaving 28K tokens for conversation (on a 128K context window model). A Minimal mode sub-agent uses only 10K tokens, leaving 118K for the actual task. This is why choosing the right mode matters for performance and cost.

Common Pitfalls

  • Using Full mode for all agents: Sub-agents that only need to perform a specific task waste tokens and may get confused by irrelevant instructions from AGENTS.md or SOUL.md.
  • Using None mode without providing custom instructions: An agent in None mode only knows its name. Without additional instructions from the parent agent's task prompt, it has no idea what to do.
  • Forgetting safety guidelines are excluded in None mode: Agents in None mode do not receive OpenClaw's safety section. Ensure the parent agent provides appropriate guardrails in the task description.

Best Practices

  • Default to Full mode for primary agents: Your main interactive agent should have the full context to handle any request.
  • Use Minimal mode for tool-using sub-agents: Sub-agents that need specific tools but not the full workspace context benefit from Minimal mode's balance of capability and efficiency.
  • Reserve None mode for tightly scoped tasks: Only use None mode when the parent agent can provide complete instructions in the task prompt and safety is handled by the parent.

Summary

  • OpenClaw offers three prompt modes: Full (complete system prompt), Minimal (identity + safety + selected tools), and None (identity only)
  • Full mode is for primary agents; Minimal for sub-agents needing some tools; None for highly specialized sub-agents
  • Token savings between modes are dramatic: Full uses 80-150K tokens, Minimal uses 5-15K, None uses 500-2K
  • None mode excludes safety guidelines, so the parent agent must provide appropriate guardrails
  • Mode is configured per agent in the agent definition, not globally
✓ Completed