Skills

Skills are dynamically loaded instruction packages that enhance the Agent’s specialized capabilities. Zeus employs a Progressive Disclosure mechanism that loads only metadata at startup and loads full content on demand, maximizing Token savings.

Reference: Anthropic Skills · Agent Skills

Three-Phase Progressive Loading

Skills loading is divided into three phases, progressively disclosing content:

Phase	Timing	Loaded Content	Token Cost
Discovery	Application startup	YAML frontmatter (name, description)	Very low
Activation	User request match or explicit activation	Full Markdown instruction content	Medium
Execution	Agent needs resources during execution	scripts/, references/, assets/	On demand

Representation in System Prompt

No Skills explicitly specified: System Prompt only injects Discovery summaries (<available_skills> XML); Agent can activate on demand via the skill_activate tool
Skills explicitly specified: Full content is directly injected into System Prompt as <activated_skill> XML

<!-- Discovery mode: metadata only -->
<available_skills>
<skill name="skill-creator">Guide for creating effective skills...</skill>
<skill name="chrome-extension">Control user browser for automated tasks...</skill>
</available_skills>

<!-- Activation mode: full content injected -->
<activated_skill name="skill-creator">
...full Markdown instructions...
</activated_skill>

SKILL.md Format

Each Skill is a directory containing a SKILL.md file with a YAML frontmatter header followed by Markdown instruction content.

---
name: my-skill
description: Describes when to use this Skill
version: 1.0.0
author: Zeus Team
tags:
  - category1
  - category2
triggers:
  - keyword1
  - keyword2
priority: 0
enabled: true
---

# Skill Instruction Content

Detailed AI instructions, rules, and examples...

Metadata Fields

Field	Required	Type	Description
`name`	Yes	`string`	Unique Skill identifier
`description`	Yes	`string`	Describes when to use this Skill (for Agent judgment)
`version`	No	`string`	Version number, defaults to `1.0.0`
`author`	No	`string`	Author
`tags`	No	`string[]`	Tag list for categorization and filtering
`triggers`	No	`string[]`	Trigger keywords for automatic matching against user messages
`priority`	No	`int`	Priority; higher values take precedence (default `0`)
`enabled`	No	`bool`	Whether enabled (default `true`)

Directory Structure

my-skill/
├── SKILL.md              # Main instruction file (required)
├── SKILL_ENABLED.md      # Enabled state instructions (Connector Skill specific)
├── scripts/              # Executable scripts (Phase 3, loaded on demand)
│   ├── init_skill.py
│   └── package_skill.py
├── references/           # Reference documents (Phase 3, loaded on demand)
│   ├── output-patterns.md
│   └── workflows.md
└── assets/               # Static resources

Skill Sources

Zeus supports multiple Skill sources:

Source	Enum Value	Description
Local	`local`	Local filesystem (`repository/skills/`)
Bundled	`bundled`	Pre-installed built-in
User Upload	`upload`	User-uploaded (via API or SKILL.md file)
Hub	`hub`	Skills Hub (community shared)
Remote	`remote`	Remote URL

Storage

Type	Storage Location	Description
Built-in Skills	`repository/skills/`	Deployed with code, discovered at startup
User Skills	Supabase Storage `users/{user_id}/skills/`	User-level, synced to memory via sync API

Sandbox Configuration File (.zeus.json)

When Skills execute in the sandbox, they may need access to user-configured API keys. Zeus automatically writes the keys configured in the user’s settings to /home/user/.zeus.json in the sandbox, along with /home/user/.openclaw.json for compatibility with the community Skill ecosystem.

File Structure

{
  "env": {
    "GEMINI_API_KEY": "AIzaSy...",
    "OPENAI_API_KEY": "sk-...",
    "UNIFYLLM_API_KEY": "sk-..."
  }
}

Aligned with the OpenClaw env block structure, using flat key-value format.

Reading Examples

Python:

import json
with open('/home/user/.zeus.json') as f:
    config = json.load(f)
api_key = config['env']['GEMINI_API_KEY']

Node.js:

const config = JSON.parse(require('fs').readFileSync('/home/user/.zeus.json', 'utf-8'))
const apiKey = config.env.GEMINI_API_KEY

Compatibility

File	Description
`/home/user/.zeus.json`	Zeus standard configuration file
`/home/user/.openclaw.json`	Community compatible, content identical to `.zeus.json`

Built-in Skills

chrome-extension

Field	Value
Type	Connector Skill
Description	Control user browser for automated tasks (open pages, click, fill forms, screenshot, etc.)
Triggers	open page, browser, search, click, screenshot, navigate…
Tags	`browser`, `automation`, `web`, `connector`

Browser Operator uses progressive disclosure:

Browser extension not connected: Loads SKILL.md (guides user to install/connect extension)
Browser extension connected: Loads SKILL_ENABLED.md (complete browser tool documentation)

desktop-operator

Field	Value
Type	Connector Skill
Description	Execute Shell commands on user’s local machine (including git, npm, file operations, etc.)
Triggers	desktop, terminal, shell, git, npm, command…
Tags	`desktop`, `shell`, `automation`, `connector`

Desktop Operator also uses progressive disclosure, switching between SKILL.md / SKILL_ENABLED.md based on desktop application connection status.

skill-creator

Field	Value
Type	Development Skill
Description	Guide creation of new Skills, providing best practices and templates
Triggers	create skill, new skill…
Tags	`skill`, `creator`, `development`, `agent`
Resources	`scripts/init_skill.py`, `scripts/package_skill.py`, `references/output-patterns.md`, `references/workflows.md`

Skill Tools (Agent-Side Tools)

The Agent can operate Skills on demand at runtime through the following tools:

Tool	Description
`skill_discover(tag?)`	List all available Skills, with optional tag filtering (Phase 1: Discovery)
`skill_activate(skill_name)`	Activate a specified Skill, loading its full content (Phase 2: Activation)
`skill_execute(skill_name, resource_type, resource_name)`	Load a Skill’s scripts/reference documents (Phase 3: Execution)

Typical Flow:

Skill Activation Configuration

The frontend can configure Skills activation via SkillActivation:

{
  "skills": {
    "skill_names": ["chrome-extension"],
    "auto_match": true,
    "max_skills": 3
  }
}

Field	Type	Description
`skill_names`	`string[]`	List of explicitly activated Skill names
`auto_match`	`bool`	Whether to automatically match based on user messages (default `true`)
`max_skills`	`int`	Maximum number of activated Skills (default `3`)

Matching Priority: Explicitly specified > Auto-matched (sorted by priority descending)

Skills Management API

Discovery Phase

Method	Path	Description
`GET`	`/api/skills`	Get all Skill metadata (supports `enabled_only`, `tags` filtering)
`GET`	`/api/skills/stats`	Get statistics (discovered, activated, enabled)
`GET`	`/api/skills/user`	Get user’s Skills in Supabase Storage

Activation Phase

Method	Path	Description
`GET`	`/api/skills/{name}`	Get Skill details (loads full content)
`POST`	`/api/skills/{name}/activate`	Explicitly activate a Skill

Execution Phase

Method	Path	Description
`GET`	`/api/skills/{name}/resources/{type}/{file}`	Get Skill resource (script/reference/asset)

CRUD

Method	Path	Description
`POST`	`/api/skills`	Create Skill (JSON request body)
`POST`	`/api/skills/upload`	Upload SKILL.md file to create Skill
`PUT`	`/api/skills/{name}`	Update Skill
`DELETE`	`/api/skills/{name}`	Delete Skill
`POST`	`/api/skills/{name}/toggle`	Toggle enabled/disabled state
`POST`	`/api/skills/reload`	Reload from directory

Storage Sync

Method	Path	Description
`POST`	`/api/skills/sync`	Sync a single Skill (Supabase → memory cache)
`POST`	`/api/skills/sync-all`	Sync all user Skills

Connector Skills Progressive Disclosure

Browser Operator and Desktop Operator are special Connector Skills that dynamically switch prompt content based on node connection status: This mechanism ensures:

The Agent will not call browser/desktop tools when no nodes are available
The Agent knows how to guide users to connect missing nodes
When connected, the Agent receives complete tool usage documentation

Skills vs MCP Prompts

Feature	Skills	MCP Prompts
Source	Local files / Supabase Storage / Hub	MCP Server
Loading	Progressive three-phase (Discovery → Activation → Execution)	Loaded at startup
Triggering	Keyword auto-match or explicit activation	Explicit invocation
Parameterization	No dynamic parameters	Supports `arguments`
Resources	Supports scripts/, references/, assets/	Not supported
Management	Filesystem + REST API + Supabase Storage	Via MCP protocol
Purpose	Enhance Agent specialized capabilities and behavior patterns	External system prompt templates

Both can be used simultaneously and complement each other. Skills focus on Agent behavioral instructions, while MCP Prompts focus on prompt templates for external system integration.

Get Started

Fundamentals

Core Capabilities

Infra

Channels

Three-Phase Progressive Loading

Representation in System Prompt

SKILL.md Format

Metadata Fields

Directory Structure

Skill Sources

Storage

Sandbox Configuration File (.zeus.json)

File Structure

Reading Examples

Compatibility

Built-in Skills

chrome-extension

desktop-operator

skill-creator

Skill Tools (Agent-Side Tools)

Skill Activation Configuration

Skills Management API

Discovery Phase

Activation Phase

Execution Phase

CRUD

Storage Sync

Connector Skills Progressive Disclosure

Skills vs MCP Prompts

​Three-Phase Progressive Loading

​Representation in System Prompt

​SKILL.md Format

​Metadata Fields

​Directory Structure

​Skill Sources

​Storage

​Sandbox Configuration File (.zeus.json)

​File Structure

​Reading Examples

​Compatibility

​Built-in Skills

​chrome-extension

​desktop-operator

​skill-creator

​Skill Tools (Agent-Side Tools)

​Skill Activation Configuration

​Skills Management API

​Discovery Phase

​Activation Phase

​Execution Phase

​CRUD

​Storage Sync

​Connector Skills Progressive Disclosure

​Skills vs MCP Prompts

Three-Phase Progressive Loading

Representation in System Prompt

SKILL.md Format

Metadata Fields

Directory Structure

Skill Sources

Storage

Sandbox Configuration File (.zeus.json)

File Structure

Reading Examples

Compatibility

Built-in Skills

chrome-extension

desktop-operator

skill-creator

Skill Tools (Agent-Side Tools)

Skill Activation Configuration

Skills Management API

Discovery Phase

Activation Phase

Execution Phase

CRUD

Storage Sync

Connector Skills Progressive Disclosure

Skills vs MCP Prompts