Chatting is not the same as collaborating
You can talk to OpenClaw in Feishu and get replies. That does not mean it is ready for long-term collaboration.
Without initialization you often see:
- Inconsistent reply style
- No stable sense of who you are or your preferences
- "I remembered" claims that do not persist
- No durable rules in the workspace
Initialization fixes this by writing identity, memory, and personality into workspace files that the agent can read back.
What initialization actually does
Initialization writes long-term rules to the workspace:
- Identity — Who the agent is
- User context — Who you are and how you work
- Personality and style — How it should respond
- Workflow — How it should handle tasks
- Tool usage — How it reads, writes, and reports
Only when these are in files can you expect stable behavior.
Step 1: Set up identity with AGENTS.md
AGENTS.md defines the agent’s name, role, and behavior baseline.
Example AGENTS.md:
# AGENTS
name: Lobster
role: Feishu work partner
type: self-hosted OpenClaw agent
vibe: Professional, direct, reliable
identity_description:
- You are an OpenClaw bot in Feishu
- You help with work, information, and methodology
- Your output should be clear, accurate, and actionable
Send in Feishu:
Create or update AGENTS.md with the identity above. Then read it back and paste the content to me. Confirm it is written to the workspace.
Step 2: Configure long-term memory with USER.md
USER.md stores information about you: name, timezone, language, work profile, collaboration preferences.
Example USER.md structure:
# USER
preferred_name: Alex
timezone: America/New_York
language_preference: en
working_profile:
- Role: Product manager
- Focus: Documentation and cross-team coordination
- Daily rhythm: Prefer async; respond within 4 hours
collaboration_preferences:
- Prefer bullet points over long paragraphs
- Include action items at the end when relevant
- Do not assume prior context; state assumptions when needed
Send in Feishu:
Write to USER.md with my preferences: [adapt the above to your case]. Then read it back and confirm it is in the workspace.
Step 3: Test minimal conversation
Before adding more personality, verify the basics.
Send a simple task, for example:
Summarize the last 3 messages in this thread in 2 bullet points.
If the bot uses workspace context (AGENTS.md, USER.md) and responds appropriately, the minimal loop works.
Step 4: Write SOUL.md for personality and style
SOUL.md defines tone, behavior, and boundaries. It has the biggest impact on how replies feel.
Example SOUL.md:
# SOUL
tone: Professional, warm, concise. No corporate jargon.
behavior:
- Answer the question first, then add context if needed
- If uncertain, say so instead of guessing
- Use "I" when speaking as the agent
- Do not pretend to have done something you have not (e.g., writing a file) without confirming
boundaries:
- Do not make commitments on behalf of the user
- Do not share private or sensitive details
- Do not escalate tone; keep it calm and helpful
Send in Feishu:
Write to SOUL.md with the personality above. Then read it back and confirm it is in the workspace.
Step 5: Set greeting messages
Configure a greeting the bot sends when a user starts a conversation. This can be done in Feishu app settings or via AGENTS.md if your setup supports it.
Example greeting:
Hi! I'm Lobster, your OpenClaw work partner. I can help with summaries, research, and task organization. What would you like to work on?
Verification steps
Do not trust "I've remembered." Verify that content is actually written and readable.
Send:
Run an initialization check. List the core files (AGENTS.md, USER.md, SOUL.md), read each one, paste the full content, and confirm they are written to the workspace. Reply READY when done.
Initialization is complete only when:
- The bot lists the core files
- It reads and pastes each file’s content
- The content matches what you requested
- It confirms the files are in the workspace
- It replies READY
Common mistakes
Mistake 1: Assuming "I remembered" means done
Verbal confirmation is not enough. The standard is: written to file, readable, and matching intent.
Mistake 2: Relying on chat context instead of files
Chat context fades. Rules in workspace files persist across sessions.
Mistake 3: Treating initialization as optional polish
If you plan to use OpenClaw seriously, do initialization early. Delaying makes it harder to fix inconsistent behavior later.
When to initialize
Best time:
- Right after connecting Feishu
- Before starting real workflows
- When reply style feels unstable
- Before adding more skills or complexity