12 Advanced OpenClaw Configurations to Train Your AI Like an Employee

SOUL File – AI personality

SOUL.md defines the AI’s core positioning, behavior rules, communication style, and prohibited actions. Example content includes concise answering, uncertainty disclosure, privacy protection, formal yet approachable tone, and bans on fabricating information. The file is placed at /workspace/SOUL.md.

Identity – Role definition

IDENTITY.md specifies the AI’s name, role, speaking style, strengths, and weaknesses. Example values: name 小墨, role AI 写作助手, style of direct, professional, example‑driven speech, strengths in article polishing and title optimization, weaknesses in math and real‑time trends. The file resides at /workspace/IDENTITY.md.

Users – User profile

USER.md records basic information, preferences, taboos, and work habits for a given user. Example entries: name 王总, industry 科技创业, preference for concise answers, dislike of long messages, morning focus on important tasks, and weekend inactivity. Stored at /workspace/USER.md.

Agents – Multi‑agent architecture

AGENTS.md describes a team of sub‑agents, each handling a domain:

Writing Agent – duties: article drafting, polishing, rewriting; skills: markdown, public‑account format; trigger words: “写文章”, “润色”.

Research Agent – duties: information search, content scraping; skills: web search, page analysis; trigger words: “查一下”, “调研”.

Coding Agent – duties: code writing, debugging; skills: Python, JavaScript; trigger words: “写代码”, “调试”.

The main scheduler routes tasks to the appropriate sub‑agent. File location: /workspace/AGENTS.md.

Tools – Available utilities

TOOLS.md lists core tools and usage constraints:

Browser Control – automate web forms, screenshots, batch operations.

File Management – read, write, organize files for reports.

Search – fetch online information.

Scheduled Tasks – daily reminders, periodic pushes.

Usage rules require user confirmation for browser actions, automatic backups for file operations, source citation for search results, and explicit timing for scheduled tasks. Stored at /workspace/TOOLS.md.

Memory enhancement – Prevent forgetting

Memory is organized in three layers:

Short‑term session memory – retains important facts within the current conversation.

Daily diary – files under memory/ (e.g., 2026-03-01.md) capture day‑by‑day records.

Long‑term memory – MEMORY.md aggregates critical user info, preferences, and configured skills.

Reading priority: current session → today’s diary → historical diary → MEMORY.md. Example snippets show user goals, tool choices, and prohibited content.

Errors – Error log

ERRORS.md logs resolved and pending issues with problem description, cause analysis, fix steps, and lessons learned. Sample resolved entries include:

WeChat link format error – missing ?from=xxx parameter; fixed by using card link with title, description, and cover.

Python dependency install failure – environment locked; fixed by creating a virtual environment with python3 -m venv.

File path error – mixed relative/absolute paths; fixed by standardizing on absolute paths.

Pending entry: browser cannot start, symptom “Chrome not found”, status “investigating”, attempted fix “install Chromium”.

Permissions – Autonomous scope

PERMISSIONS.md separates actions that the AI may perform without confirmation from those that require explicit approval:

Can do directly : network information retrieval, reading existing files, calendar lookup, text polishing, typo correction, format tweaks, simple code fixes, creating new files, sending messages to configured channels, taking screenshots.

Requires confirmation : deleting files, modifying configuration files, sending external email, executing system commands, accessing unauthorized directories, using payment functions, modifying user data.

File location: /workspace/PERMISSIONS.md.

Skill loading – Extensible capabilities

Skills are packaged in a directory with a mandatory SKILL.md that defines name, description, trigger, and usage steps. Example skill “公众号排版” converts markdown to WeChat HTML, offers 13 templates, and pushes the result to a draft box. Installation commands: clawhub install 技能名称 or manual copy: cp -r my-skill /workspace/skills/ Skill directory structure includes README.md, scripts/, and assets/.

Multi‑instance – Parallel execution

Sub‑agents can be spawned concurrently using asynchronous gathering. Example Python snippet:

agents = [
    spawn(agent="写作", task="写AI文章"),
    spawn(agent="写作", task="写理财文章"),
    spawn(agent="写作", task="写成长文章")
]
results = await gather(agents)

This enables simultaneous article generation, research, or multi‑platform publishing.

Security – Limited authorization

SECURITY.md defines access control and operation limits:

Allowed directories : /workspace/*, /tmp/*, configured external APIs.

Disallowed directories : ~/.ssh/, ~/.aws/, /etc/.

Prohibited commands : rm -rf /, curl | bash, real‑money transaction commands.

Audit : logs timestamp, action, and result of each sensitive operation.

File location: /workspace/SECURITY.md.

Daily training – Continuous improvement

PLAYBOOKS/ stores daily teaching records. Each entry includes:

What was taught (e.g., title formula, interaction prompts).

What succeeded (e.g., engaging titles, automatic paragraphing).

What needs improvement (e.g., verbosity, missing interaction prompts).

Focus for the next day (e.g., professional tech article writing, leveraging hot topics).

Example file tree:

/workspace/
├── SOUL.md
├── USER.md
├── AGENTS.md
├── TOOLS.md
├── ERRORS.md
└── PLAYBOOKS/
    ├── 写作规范.md
    ├── 标题公式.md
    └── 选题方向.md

Overall workflow

By configuring the twelve components—SOUL.md, IDENTITY.md, USER.md, AGENTS.md, TOOLS.md, the memory hierarchy, ERRORS.md, PERMISSIONS.md, skill packages, multi‑instance scripts, SECURITY.md, and PLAYBOOKS/—OpenClaw becomes a self‑learning, task‑oriented AI assistant capable of writing, research, coding, and multi‑platform publishing with minimal supervision.