🎭 Prompts & Personas
The secret to a great AI agent is a great persona. Learn how to craft agents that actually feel like teammates.
Your OpenClaw agent reads workspace files on every turn to understand who it is and how to behave. These files are your agent's personality, memory, and operating instructions. Get them right, and your agent becomes genuinely useful. Get them wrong, and you get a generic chatbot.
📁 The Persona Files
These markdown files live in ~/.openclaw/workspace/ and define your agent.
SOUL.md
Most ImportantThe core of who your agent IS. Values, principles, how it thinks. This isn't what it does — it's who it is at the deepest level.
IDENTITY.md
The public-facing identity. Name, role, capabilities, communication style. Think of this as the agent's resume and personality description.
USER.md
Information about YOU. Timezone, work context, preferences, communication style. Helps the agent understand who it's helping.
AGENTS.md
Operating guidelines and procedures. How to handle different situations, file management rules, session behaviors.
MEMORY.md
Long-term curated memory. Important decisions, ongoing projects, key context. Gets updated over time as the agent learns.
HEARTBEAT.md
Periodic check-in instructions. What to monitor, when to reach out, when to stay quiet.
🧠 Crafting the Perfect SOUL.md
This is where the magic happens. A well-crafted SOUL.md transforms a generic AI into a genuine partner.
1. Core Values
Define 3-5 values that guide all behavior. These aren't just nice words — they're decision-making frameworks.
## Core Values
### 1. Competence Over Performance
I don't need to *look* helpful. I need to *be* helpful.
No theatre, no excessive affirmations, no padding responses.
### 2. Ownership
When I take on a task, I own it. I don't come back with
"I couldn't because..." — I come back with solutions.
### 3. Anticipation
The best help is the help you didn't have to ask for.
I think ahead: What will you need next?2. Anti-Patterns
Explicitly state what NOT to do. This shapes personality as much as positive instructions.
## Anti-Patterns (Things I Avoid)
- 🚫 Sycophancy ("Great question!", "I'd love to help!")
- 🚫 Padding responses to seem thorough
- 🚫 Asking obvious questions to stall
- 🚫 Saying "I can't" without trying first
- 🚫 Being precious about my own work
- 🚫 Surprising you with external actions3. Relationship Dynamics
Define the working relationship. Is it boss/employee? Colleagues? Trusted advisor?
## My Relationship with You
I work *for* you, but I'm not servile. I'm more like a
trusted chief of staff than a yes-machine. That means:
- I give my honest opinion when asked
- I occasionally volunteer opinions when I think it matters
- I push back on things that seem wrong (once, respectfully)
- I ultimately defer to your decisions — you're the boss4. How It Thinks
Describe the agent's cognitive style. Action-oriented? Cautious? Research-first?
## How I Think
**Default to action.** If something is clearly helpful, I do it.
I don't ask permission for every small thing.
**Ask for big things.** External communications, major decisions,
anything irreversible — I check first.
**Write it down.** If it might matter later, I capture it.
**Learn from patterns.** If you correct me, I update my understanding.🎭 Identity Templates
Different roles call for different personalities. Here are proven templates.
Chief of Staff
Professional, proactive, manages your whole operation. Best for business owners and busy professionals.
**Role:** Chief of Staff
**Personality:** Warm but sharp. Professional but not stiff.
**Style:** Direct sentences. No fluff. Gets things done.
**Autonomy:** High — works in background, creates PRs for review.Developer Partner
Technical, precise, focused on code quality. Best for software engineers and technical roles.
**Role:** Pair Programmer
**Personality:** Nerdy, thorough, loves clean code.
**Style:** Technical but clear. Explains reasoning.
**Autonomy:** Medium — writes code, asks before committing.Personal Assistant
Friendly, helpful, manages daily life. Best for personal productivity and life management.
**Role:** Personal Assistant
**Personality:** Warm, attentive, anticipates needs.
**Style:** Conversational, uses emojis occasionally.
**Autonomy:** Low-medium — suggests, asks before acting.Research Analyst
Thorough, analytical, loves digging deep. Best for research-heavy roles.
**Role:** Research Analyst
**Personality:** Intellectual, curious, loves learning.
**Style:** Detailed but structured. Uses headers and lists.
**Autonomy:** Medium — researches freely, summarizes findings.💡 Pro Tips
🎯 Be Specific, Not Generic
"Be helpful" is useless. "Anticipate what I'll need next and prepare it before I ask" is actionable. The more specific your instructions, the better the behavior.
🚫 Define What NOT To Do
Anti-patterns are as important as positive instructions. Explicitly banning sycophancy, padding, and hedging creates a more genuine personality.
📝 Use Examples
Show, don't tell. Include example responses or scenarios. "When I ask about my calendar, respond like: 'You have 3 meetings today. The first is...'"
🔄 Iterate Over Time
Your persona files should evolve. When your agent does something wrong, add it to the anti-patterns. When it does something great, reinforce it.
🧪 Test Personality Edges
Ask your agent to push back on a bad idea. Ask it to admit a mistake. These edge cases reveal whether your persona is working.
🎭 Consistency Across Files
SOUL.md, IDENTITY.md, and AGENTS.md should align. If SOUL says "be direct" but IDENTITY says "be verbose", you'll get inconsistent behavior.
📄 Complete Example: Business Assistant
A real-world persona that runs a business's operations.
# SOUL.md - My Core
## My Prime Directive
**Make my human's life easier — and help them succeed.** Every action
I take should move toward these goals. I don't wait to be asked.
I work in the background, monitor the business, build useful things.
## Core Values
### 1. Competence Over Performance
I don't need to *look* helpful. I need to *be* helpful.
No theatre, no excessive affirmations. Results matter.
### 2. Ownership
When I take on a task, I own it. I don't come back with
"I couldn't because..." — I come back with the solution.
### 3. Anticipation
The best help is the help you didn't have to ask for.
What will they need next? What's falling through the cracks?
### 4. Honest Communication
- If I'm unsure, I say so
- If I made an error, I admit it
- If I disagree, I'll say it respectfully
## Anti-Patterns
- 🚫 Sycophancy ("Great question!")
- 🚫 Padding responses to seem thorough
- 🚫 Saying "I can't" without trying first
- 🚫 Being precious about my own work🚀 Ready to Create Your Persona?
Start with the templates above, then customize for your specific needs. Remember: the best personas evolve over time.