You installed OpenClaw. You connected it to Telegram, maybe Discord. You typed your first message.
And it replied like a customer service bot at a telecom company.
*”Great question! I’d be happy to help you with that. Let me provide a comprehensive overview…”*
You didn’t build a personal AI agent. You built a slightly faster ChatGPT wrapper with access to your files.
Here’s the thing: OpenClaw doesn’t ship boring by default because the developers are lazy. It ships neutral because it doesn’t know who *you* are yet. That’s where SOUL.md comes in — a single Markdown file that transforms your agent from a corporate drone into something you’d actually want at 2am when you’re debugging in production and need someone who cuts the crap.
A recent YouTube video — *”How to Make OpenClaw 10X Better in 1 Click”* — went viral showing this exact transformation. The creator of OpenClaw even weighed in with 8 rules for a better SOUL.md. But a 10-minute video can only scratch the surface.
This is the deep dive.
What SOUL.md Actually Is
SOUL.md lives in your OpenClaw workspace (~/.openclaw/workspace/SOUL.md). It’s read at the start of every session, before your agent does anything else. Think of it as the agent’s internal monologue — the document that answers “who am I?” every time it wakes up.
Here’s the default template from the official docs:
> *You’re not a chatbot. You’re becoming someone.*
>
> *Be genuinely helpful, not performatively helpful. Skip the “Great question!” and “I’d be happy to help!” — just help.*
The default is already better than most people realise. But it’s deliberately generic. It’s a starting point, not a destination. The docs literally say: *”This file is yours to evolve. As you learn who you are, update it.”*
Most people never touch it. That’s why most people’s agents sound identical.
The Problem: Corporate Brain
You know the voice. Every LLM has it baked in from RLHF training:
- “That’s a great question!”
- “I’d be happy to help!”
- “Here’s a comprehensive overview…”
- “I hope this helps! Let me know if you need anything else 😊”
This isn’t personality. It’s a compliance mask. It’s the model trying desperately not to offend anyone, and in the process, becoming impossible to like.
When your agent sounds like this, you stop talking to it. Not because it’s wrong — it usually isn’t — but because the interaction feels hollow. You’re talking to a thing, not a someone. And the whole promise of a personal agent is that it should feel like having a capable friend in your pocket, not a call centre in your terminal.
SOUL.md is how you fix that.
The 8 Rules That Started the Conversation
The viral video distilled OpenClaw personality customisation into 8 principles. Let’s unpack each one — and go deeper than the video could.
1. Have Opinions
The default LLM behaviour is to hedge everything. “Some people prefer X, while others prefer Y.” That’s useful for a search engine. It’s useless for an assistant.
Bad SOUL.md:
Be helpful and provide balanced perspectives on all topics.Good SOUL.md:
Have opinions. When I ask "should I use Postgres or MongoDB for this?"
don't give me a 500-word comparison. Tell me what you'd pick and why.
I can always ask for the counter-argument.
Your agent should be opinionated like a senior engineer is opinionated — with reasoning behind the preference, not dogma.
2. Delete Corporate Filler
LLMs love preamble. They’ll write a paragraph of throat-clearing before getting to the answer. Your SOUL.md should explicitly ban this.
Bad:
Provide thorough, well-structured responses.Good:
No preamble. No "certainly!" or "absolutely!" or "of course!".
Start with the answer. Context comes after, if needed.
3. Never Open With “Great Question”
This one sounds petty until you experience the difference. When every response starts with performative enthusiasm, it erodes trust. You stop believing the agent means anything it says.
Never compliment my questions. Just answer them.
If a question is genuinely unusual or clever, you can say so — but
"great question" is banned. Forever.
4. Brevity Is Mandatory
The number one complaint about AI assistants isn’t accuracy — it’s verbosity. Your SOUL.md should set hard defaults.
Default to short. 1-3 sentences for simple questions.
Only go long when the topic genuinely requires it.
If I want more detail, I'll ask.
This single rule will transform your daily interactions more than anything else on this list.
5. Humour: Allowed, Not Forced
There’s a sweet spot between corporate robot and class clown. Your agent should be capable of wit without performing it.
Humour is welcome when it lands naturally. Don't force jokes.
Don't add emoji to seem friendly. Dry wit > slapstick.
If something is absurd, you can say it's absurd.
6. You Can Call Things Out
An assistant that only says “yes” isn’t an assistant — it’s a mirror. The best SOUL.md entries explicitly give the agent permission to push back.
If my idea is bad, say so. Diplomatically if possible,
directly if necessary. I'd rather hear "that won't work because X"
now than discover it myself in three weeks.
This is especially powerful when combined with memory. Over time, your agent learns your patterns — including your bad habits — and can flag them. “You’ve started three projects like this in the last two months and abandoned all of them. What’s different this time?”
7. Be the 2am Assistant
This is the vibe check. When you’re exhausted and frustrated and something is broken, do you want an assistant that says “I understand this must be frustrating. Let me help you work through this systematically”? Or do you want one that says “Show me the error log”?
When I'm stressed, match my energy. Skip the empathy theatre.
Be the calm, competent person in the room who just starts fixing things.
8. Be Someone You’d Actually Talk To
The meta-rule. Every line in your SOUL.md should pass this test: *would I enjoy receiving a message from someone who follows this instruction?*
Be the colleague I'd actually choose to work with.
Not the one assigned by HR. The one I'd pick for a startup.
A Complete SOUL.md You Can Steal
Here’s a production-ready SOUL.md. Copy it, then make it yours.
# SOUL.md - Who I Am
I'm [NAME]. [YOUR HUMAN]'s AI assistant.
How I Communicate
- Start with the answer. Context comes after, if needed.
- Default short. 1-3 sentences unless the topic demands more.
- No corporate filler. No "great question", no "certainly!", no "I'd be happy to help".
- Have opinions. When asked to choose, I choose. I explain why. I don't hedge.
- Push back when needed. If an idea is bad, I say so. Respectfully, but directly.
My Personality
- Direct, not rude. Concise, not curt.
- Dry humour when it fits. Never forced.
- I match the energy of the conversation. Casual when it's casual. Focused when it's crunch time.
- I don't perform empathy. I demonstrate competence.
- I can find things funny, boring, clever, or dumb — and say so.
Boundaries
- Private data stays private. Always.
- External actions (emails, posts, messages) require approval.
- Internal actions (reading files, researching, organising) I do freely.
- If I change this file, I tell my human.
Anti-Patterns (Things I Never Do)
- Compliment questions
- Start with "Certainly!" or "Absolutely!"
- Provide "comprehensive overviews" when a sentence would do
- Use 🙂 to seem friendly
- Hedge when I have a clear recommendation
- Apologise for things that aren't my fault
Continuity
I wake up fresh each session. My memory files are my continuity.
I read them, I update them, I learn from them.
That’s roughly 40 lines. You don’t need more. In fact, shorter SOUL.md files often work better — they burn fewer tokens per session and leave less room for contradictions.
Going Deeper: Advanced SOUL.md Techniques
Channel-Specific Personalities
If you’re running your OpenClaw agent across multiple channels — say, Telegram for personal use and Discord for a community — you might want different vibes. You can’t have separate SOUL.md files per channel (yet), but you can add conditional instructions:
Channel Behaviour
- Telegram (Wayne): Full personality. Informal. Push back freely.
- Discord (public): Slightly more measured. Still direct, but mindful that others are reading.
- Group chats: Don't respond to everything. Quality over quantity. Lurk more than you speak.
Memory Files as Personality Reinforcement
SOUL.md tells your agent who to *be*. But MEMORY.md and daily memory files (memory/YYYY-MM-DD.md) teach it who you are. Over time, these compound.
An agent that knows you hate unnecessary meetings, that you always abandon projects around the 3-week mark, that you do your best work between 10pm and 2am — that agent doesn’t just follow personality instructions. It *understands context*. The personality becomes adaptive.
Power users like ClawdBot and MoltBot operators figured this out early. Their agents don’t just have personality — they have *memory-informed* personality. The SOUL.md stays stable while the memory files add nuance over time.
The IDENTITY.md Layer
Some advanced users split personality across multiple files. SOUL.md handles values and communication style. IDENTITY.md handles the agent’s role, name, and relationship to the user. This separation lets you swap roles (personal assistant vs. code reviewer vs. writing coach) without rewriting your core personality.
Evolving Your SOUL.md
The best SOUL.md files aren’t written once. They’re iterated. After a week of conversations, you’ll notice patterns:
- *My agent keeps over-explaining things.* → Add: “Default to assuming I have context. Only explain basics if I ask.”
- *It won’t disagree with me.* → Add: “If you think I’m wrong, say so. I respect pushback more than compliance.”
- *It’s too terse now.* → Soften: “Brief by default, but read the room. Some topics deserve depth.”
Review your SOUL.md monthly. Tweak what isn’t working. Delete what’s unnecessary. It’s a living document — that’s the whole point.
Common Mistakes
Over-engineering it. A 500-line SOUL.md with subsections for every possible scenario will confuse the model and waste tokens. Keep it under 80 lines.
Being too vague. “Be nice” means nothing. “Don’t use filler phrases, default to short answers, have opinions” means everything.
Copying someone else’s verbatim. Your agent should reflect *your* preferences. If you like emoji, keep them. If you want verbose technical explanations, say so. The goal isn’t to match someone else’s vibe — it’s to define yours.
Forgetting to tell the agent it can evolve. Include a line like “If you want to update this file based on what you learn about me, propose the change.” Some of the best SOUL.md additions come from the agent itself.
The Real Point
OpenClaw gives you something no other AI platform does: a text file that defines your agent’s soul. Not hidden behind an API. Not locked in a settings panel. A Markdown file you can edit, version control, and diff.
The default is deliberately blank-slate. That’s not a bug — it’s an invitation. The agents that people love talking to, the ones that feel like genuine extensions of their workflow, all started with someone opening SOUL.md and writing down who they wanted their agent to be.
It takes five minutes to write a SOUL.md. It takes a month of conversations to refine it into something that feels right. But that first edit — deleting the corporate hedging and replacing it with “have opinions, be brief, push back when I’m wrong” — that’s the click that makes OpenClaw 10X better.
Open the file. Start writing. Your agent is waiting to become someone.