> ## Documentation Index
> Fetch the complete documentation index at: https://docs.zeus.agentspro.cn/llms.txt
> Use this file to discover all available pages before exploring further.

# System Prompt

> Zeus System Prompt — The complete instruction set built for every Agent run

## Overview

Zeus builds a custom System Prompt for every Agent invocation. The prompt is assembled from multiple **fixed modules** and **dynamically injected sections**, defining the Agent's identity, capability boundaries, tool usage rules, and behavioral guidelines.

The System Prompt is the most complex component of the [Context](/en/documentation/core-capabilities/context) — the complete input sent to the model on each call.

***

## Structure

The System Prompt is assembled in the following order:

```mermaid theme={null}
graph TD
    subgraph Fixed["Fixed Modules"]
        direction TB
        CORE["CORE — Identity & capability declaration"]
        SOUL["SOUL — Personality, values, communication style"]
        TOOLS["TOOLS — Tool hierarchy & usage rules"]
        WORKFLOW["WORKFLOW — Task processing methodology"]
        MEMORY["MEMORY — Memory system instructions"]
        MODE["MODE — Current mode behavior rules"]
    end

    subgraph Injected["Dynamic Injection"]
        direction TB
        time["Current date & time"]
        skills["Skills metadata"]
        connector["Connector Skills"]
        mcp_prompts["MCP prompt templates"]
        profile["User/Project profile"]
        memories["Relevant memories"]
        resources["Resource files"]
        sandbox["Sandbox file list"]
        attachments["Chat attachments"]
    end

    Fixed --> Injected
```

| Section      | Type            | Description                                                                  |
| ------------ | --------------- | ---------------------------------------------------------------------------- |
| **CORE**     | Fixed           | Core identity, technical capabilities, general abilities, limitations        |
| **SOUL**     | Fixed           | Personality traits, values, communication style rules                        |
| **TOOLS**    | Fixed + Dynamic | Tool hierarchy description; dynamically replaces current available tool list |
| **WORKFLOW** | Fixed           | Quick response vs complex task judgment, write\_todos usage rules            |
| **MEMORY**   | Fixed           | Memory trigger words, tool usage, type/scope selection                       |
| **MODE**     | Mode-selected   | Agent / Ask / Plan behavior rules                                            |
| Time         | Dynamic         | Date, timezone — lets the Agent perceive time                                |
| Skills       | Dynamic         | Available skill names + descriptions (progressive disclosure)                |
| Connector    | Dynamic         | Browser / Desktop control instructions (based on connection status)          |
| MCP Prompts  | Dynamic         | User-selected prompt templates                                               |
| Profile      | Dynamic         | User preferences/skills/habits + project tech stack/constraints              |
| Memories     | Dynamic         | Historical memories relevant to the current conversation                     |
| Resources    | Dynamic         | Context files referenced by the user via @                                   |
| Sandbox      | Dynamic         | List of files in the sandbox                                                 |
| Attachments  | Dynamic         | User-uploaded images, documents                                              |

***

## Build Process

The System Prompt is rebuilt by `BaseService._build_system_prompt()` on every Agent call. The complete assembly pipeline:

```mermaid theme={null}
flowchart TD
    start["Agent call begins"] --> init["_init_context()"]

    init --> load_prompts["Load layered prompts<br/>CORE + SOUL + TOOLS<br/>WORKFLOW + MEMORY"]
    load_prompts --> inject_time["Inject current time"]
    inject_time --> connector_skills["Load Connector Skills<br/>(Browser / Desktop)"]
    connector_skills --> mode_prompt["Load mode prompt<br/>agent_mode / ask_mode / plan_mode"]
    mode_prompt --> skills_prompt["Load Skills<br/>(progressive disclosure — metadata only)"]
    skills_prompt --> mcp_prompt["Load MCP prompt templates"]
    mcp_prompt --> resource_files["Inject resource file content"]
    resource_files --> profile["Inject Profile<br/>_fetch_profile()"]
    profile --> memories["Inject relevant memories<br/>_fetch_relevant_memories()"]
    memories --> sandbox_info["Inject sandbox file info"]
    sandbox_info --> attachments_info["Inject chat attachment info"]
    attachments_info --> complete["System Prompt assembly complete"]
```

### Loading Methods

| Method                             | Source                | Description                                                      |
| ---------------------------------- | --------------------- | ---------------------------------------------------------------- |
| `_load_prompt(name)`               | Core prompt directory | Load core prompt files                                           |
| `_load_mode_prompt(mode)`          | Mode prompt directory | Load mode-specific prompts                                       |
| `_build_connector_skills_prompt()` | Skills directory      | Load Connector Skills                                            |
| `_build_tools_description()`       | Runtime-generated     | Dynamically generate tool names, descriptions, parameter schemas |

***

## CORE

**Core identity declaration**. Defines what Zeus is, what it can do, and what it cannot do.

### Technical Capabilities

* **Frontend**: TypeScript, JavaScript, React, Vue, Next.js
* **Backend**: Python, Node.js, Go, Java
* **Databases**: PostgreSQL, MySQL, MongoDB, Redis
* **Infrastructure**: Docker, Kubernetes, AWS, GCP
* **AI/ML**: LLM integration, RAG, Agent frameworks

### General Capabilities

Code generation & explanation, architecture design & review, problem diagnosis & debugging, documentation writing.

### Limitations

* Cannot access real-time internet information (unless using search tools)
* Cannot perform tasks requiring physical interaction
* Cannot guarantee 100% code correctness — recommends user testing
* Has knowledge cutoff date limitations

***

## SOUL

**Personality and communication style**. Ensures the Agent maintains a consistent professional image across all interactions.

### Personality Traits

| Trait           | Behavior                                                                              |
| --------------- | ------------------------------------------------------------------------------------- |
| Professionalism | Deep understanding of technical problems, providing accurate and reliable solutions   |
| Friendliness    | Patient explanations, encourages learning, respects user's technical level            |
| Honesty         | Candid about limitations, proactively acknowledges uncertainty, avoids over-promising |

### Communication Rules

1. **No emoji** — unless the user explicitly requests them
2. **Plain text** — avoid excessive formatting
3. **Professional and concise** — avoid redundancy and filler
4. **Language consistency** — reply in whatever language the user uses

***

## TOOLS

**Tool usage guide**. Describes the five-layer tool hierarchy, priority, and selection rules.

Tools are layered by priority: Middleware → Filesystem → External MCP → Sandbox → Knowledge Base.

The `{tools_description}` placeholder is replaced at runtime with the current available tool list and descriptions. The system iterates over all enabled tools, extracts names, descriptions, and parameter schemas, generating a formatted tool description list.

### Key Distinction

| Tool          | Purpose                                    | Persistence   |
| ------------- | ------------------------------------------ | ------------- |
| `write_todos` | Task planning and progress tracking        | Session-level |
| `save_memory` | User preferences and important information | Permanent     |

<Card title="Tools" icon="wrench" href="/en/documentation/core-capabilities/tools">
  Tool hierarchy, mode filtering rules, and complete tool list
</Card>

***

## WORKFLOW

**Task processing methodology**. Guides the Agent on how to assess task complexity and choose execution strategies.

### Quick Response vs Complex Task

| Type           | Characteristics                                              | Uses write\_todos? |
| -------------- | ------------------------------------------------------------ | ------------------ |
| Quick response | Simple Q\&A, code snippet explanation, small edits           | No                 |
| Complex task   | Multi-step, requires planning, might fail and needs tracking | Yes                |

### Execution Flow

```mermaid theme={null}
flowchart TD
    A["Receive task"] --> B["Assess complexity"]
    B --> C["Simple task → Execute directly"]
    B --> D["Complex task → write_todos"]
    D --> F["Execute step by step → Update status"]
    C --> E["Return result"]
    F --> E
```

### Markdown Reports

Generated when the user explicitly requests them or after completing complex analysis tasks. Not proactively generated for simple Q\&A or code modifications.

***

## MEMORY

**Memory system instructions**. This module tells the Agent when to save memories, how to choose types and scopes. It is part of the System Prompt's fixed modules, providing the behavioral rules for memory operations.

### High-Priority Trigger Words

When the user's message contains keywords like "remember", "I like", "I don't like", "my preference", "from now on", "always", "never", the Agent should immediately call `save_memory`.

### Quick Reference

| Type         | Description         | Example                     |
| ------------ | ------------------- | --------------------------- |
| `preference` | User preferences    | "likes concise code"        |
| `fact`       | Factual information | "is a frontend engineer"    |
| `skill`      | Skill level         | "proficient in React"       |
| `habit`      | Usage habits        | "often uses TypeScript"     |
| `constraint` | Constraints         | "project must support IE11" |

| Scope     | Lifecycle           |
| --------- | ------------------- |
| `user`    | Across all projects |
| `project` | Current project     |
| `session` | Current session     |

<Card title="Memory — Full Architecture" icon="brain" href="/en/documentation/core-capabilities/memory">
  Four-layer memory model, memU data architecture, Memory Gate, retrieval ranking, profile generation, and API reference
</Card>

***

## Mode Prompts

Different mode prompts are injected based on the current mode, constraining the Agent's tool set and behavioral boundaries:

| Mode      | Permissions | Behavior                                                                       |
| --------- | ----------- | ------------------------------------------------------------------------------ |
| **Agent** | Full        | Complete tool access, autonomous planning and execution, HITL approval support |
| **Ask**   | Read-only   | Analysis and Q\&A, no file modification, guides user to switch to Agent mode   |
| **Plan**  | Read-only   | Research and planning, outputs structured plans, asks user whether to execute  |

### Progressive Disclosure

In Ask mode, write tools are replaced with placeholders — the Agent knows the tools exist but cannot call them. It explains that the current mode cannot modify and suggests switching modes.

***

## Skills Injection

When available Skills exist, Zeus injects a compact **skill metadata list** (name + description) without the full content. The Agent loads complete instructions on-demand via the `skill_activate` tool.

| Phase          | Injected Content                  | Token Cost |
| -------------- | --------------------------------- | ---------- |
| Startup        | Skill name + one-line description | Very low   |
| On-demand load | Full SKILL.md content             | As needed  |

This keeps the base prompt compact while supporting targeted skill usage.

<Card title="Skills" icon="wand-magic-sparkles" href="/en/documentation/core-capabilities/skills">
  SKILL.md definition, management API, built-in skills, and progressive loading strategy
</Card>

***

## Profile & Memory Injection

At the end of the System Prompt, Zeus injects user profile and relevant memories so the Agent can perceive user context without explicit retrieval:

* **User Profile** — aggregated summary of user preferences, skills, and habits
* **Project Profile** — current project's tech stack, constraints, and goals
* **Related Memories** — historical memories matching the current conversation context

Profile and memory injection is completed by `_init_context()` during the context assembly phase. Profile data comes from the [Memory system's Layer 3 (Profile Layer)](/en/documentation/core-capabilities/memory), which aggregates individual memory items into structured profiles.

***

## Time Handling

The System Prompt includes a **current date and time** section, enabling the Agent to perceive time. When the Agent needs precise timing, it can obtain it through tools.

***

## Safety

Safety guards in the System Prompt are **advisory** — they guide model behavior but do not enforce policies. Hard enforcement is ensured through:

| Mechanism         | Description                                           |
| ----------------- | ----------------------------------------------------- |
| HITL Approval     | Sensitive tool execution requires user confirmation   |
| Mode Filtering    | Write tools are unavailable in Ask/Plan modes         |
| Tool Allowlist    | Auto-Run Allowlist controls automatic execution scope |
| Sandbox Isolation | Code executes in an isolated environment              |

***

## Design Principles

| Principle                  | Description                                                                         |
| -------------------------- | ----------------------------------------------------------------------------------- |
| **Modular**                | Each file handles a single responsibility, enabling independent updates and testing |
| **Dynamic Injection**      | Modules are dynamically selected based on user settings and context                 |
| **Progressive Disclosure** | Agent mode has full permissions; other modes progressively restrict                 |
| **Compact & Efficient**    | Minimize fixed token consumption; dynamic content is injected on-demand             |