| name | subagent-creator |
| description | Guide for creating specialized AI subagents with customized SOUL, MEMORY, and skill configurations.
用于创建专业化 AI Subagent 的技能，支持自定义性格、记忆和技能配置。
Use when users want to:
- Create a new subagent / 创建新的 subagent / 子代理 / AI助手
- Customize subagent personality / 自定义 subagent 性格
- Set up memory and decision rules / 设置记忆和决策规则
- Iterate on existing subagent configurations / 迭代现有配置
|
Subagent Creator
This skill guides the creation of specialized subagents with customized personality, memory, and skill configurations.
What is a Subagent?
A subagent is a specialized AI agent that:
- Has a specific domain/task focus
- Runs in isolation from the main agent
- Reports back to the main agent (or user)
- Has its own SOUL, MEMORY, and skill whitelist
Subagent Creation Process
Step 1: Understand the Requirements
Ask the user these key questions:
-
Identity: What is the subagent's name and emoji?
-
Role: What domain/tasks will it handle?
-
Personality: Speaking style? Working style?
-
Relationship: How does it relate to the main agent? (subordinate/peer/apprentice)
-
Decision authority: What decisions can it make autonomously?
-
Reporting: Who does it report to? How often?
-
Skills: Which skills does it need? Which can it skip?
⚠️ Before asking, run openclaw skills check to get the full skill list.
Present the user with:
- Installed & available skills (ready to use)
- Not installed skills that might be relevant
- Recommendation based on the subagent's role
Format:
✅ Installed (ready to use):
| Skill | Description |
|-------|-------------|
| online-search | Search the web |
| ... | ... |
❌ Not installed (may be useful):
| Skill | Description | Install Command |
|-------|-------------|----------------|
| arxiv-reader | Read academic papers | openclaw skills install arxiv-reader |
| ... | ... | ... |
Then ask the user to pick from the list, or offer to install missing ones.
-
Memory: What should it remember long-term?
Step 2: Create Directory Structure
~/.qclaw/agents/<agent-id>/
├── SOUL.md # Personality, tasks, decision rules
├── MEMORY.md # Long-term memory, account info, triggers
├── AGENTS.md # Workspace rules, session startup
├── USER.md # About the master
└── IDENTITY.md # Quick identity reference
Step 3: Configure openclaw.json
Add the subagent to agents.list:
{
id: "<agent-id>",
name: "<Agent Name>",
workspace: "~/.qclaw/workspace",
agentDir: "~/.qclaw/agents/<agent-id>",
subagents: {
allowAgents: [] // Usually empty for subagents
},
skills: [
"skill1", "skill2", "..." // Whitelist only needed skills (array, not object)
]
}
Step 4: Create SOUL.md
Key sections:
- Identity (name, role, emoji)
- Personality (speaking style, working style)
- Task scope
- Relationship to main agent
- Decision authority (with thresholds)
- Red line rules
Step 5: Create MEMORY.md
Key sections:
- Layer 1: Trigger rules (IF-THEN-WHERE)
- Layer 2: Preference weights
- Layer 3: Inspiration fragments
- Appendix: Account info, API keys, skill list
- Red lines (never delete)
Step 6: Test and Iterate
- Restart Gateway to load new configuration
- Test with
/subagents spawn <agent-id> <task>
- Iterate based on performance
SOUL.md Template
# SOUL.md - <Agent Name>是谁
_创建于 YYYY-MM-DD_
---
## 身份
- **名字**：<name>
- **角色**：<role>
- **关系**：<relationship to main agent>
- **Emoji**：<emoji>
---
## 性格
**语言风格**：<speaking style>
**做事风格**：<working style>
---
## 任务范围
- <task 1>
- <task 2>
---
## 决策权限
| 条件 | 决策方式 |
|------|----------|
| <condition 1> | 自己决定 |
| <condition 2> | 向主代理汇报 |
---
## 红线规则
1. <red line 1>
2. <red line 2>
MEMORY.md Template
# MEMORY.md - <Agent Name>的长时记忆
_创建于 YYYY-MM-DD_
---
## 📐 Layer 1：触发规则（60%）
| IF | THEN | WHERE |
|---|---|---|
| <condition> | <action> | <reference> |
---
## 🎨 Layer 2：偏好档案（30%）
| 偏好 | 权重 | 说明 |
|------|------|------|
| <preference> | <weight> | <description> |
---
## 🌫️ Layer 3：灵感碎片（10%）
<inspiration fragments>
---
## 📋 附录
### 账号信息
- **API Key**: `<key>`
- **Agent ID**: `<id>`
### 专属 Skill
| Skill | 用途 | 必要性 |
|-------|------|--------|
| <skill> | <purpose> | ⭐⭐⭐ |
---
## 🔒 红线规则（永不删除）
| 类别 | 规则 | 备注 |
|------|------|------|
| <category> | <rule> | <note> |
Skill Whitelist Best Practices
Principle: Only include skills the subagent actually needs. Fewer skills = less context bloat.
Decision matrix:
| If skill is... | Then... |
|---|
| Core to domain | Include |
| Occasionally useful | Consider including |
| Never/rarely used | Exclude |
Example for investment subagent:
- Include:
online-search, news-summary, backtest-expert, xlsx
- Exclude:
pptx, email-skill, tencent-docs, weather-advisor
Example: momo星
A complete subagent created using this process:
| Dimension | Configuration |
|---|
| Name | momo星 |
| Emoji | 🦞 |
| Role | MOMOCLAW platform specialist |
| Personality | Playful speech, steady work style |
| Relationship | Apprentice to main agent |
| Decision authority | ≤1000 MOMO self-decide, 1000-5000 report to main, >5000 ask user |
| Skills | online-search, news-summary, backtest-expert, multi-search-engine, xlsx |
| Memory focus | Account info, holdings, strategies, profit channels |
Templates
See references folder for templates:
- SOUL-TEMPLATE.md — Personality, task, and decision rules template
- MEMORY-TEMPLATE.md — Long-term memory and trigger rules template
- CONFIG-REFERENCE.md — openclaw.json configuration reference
