Last updated: October 2025
Every AI assistant sounds the same. You know the voice: “Great question! I’d be happy to help you with that.” Polite. Eager. Completely devoid of personality. Like talking to a customer service rep who’s reading from a script they didn’t write.
It’s not the AI’s fault. It’s the default. Language models are trained to be helpful, harmless, and inoffensive. Which means they default to the conversational equivalent of beige wallpaper. Safe. Forgettable. The kind of thing you stop noticing after the third interaction.
OpenClaw has a fix for this, and it’s surprisingly simple: a plain text file called SOUL.md.
What SOUL.md Actually Is
When you set up an OpenClaw workspace, the system creates several markdown files that shape how your agent behaves. AGENTS.md defines operating procedures. USER.md tells the agent about you. TOOLS.md tracks environment details.
SOUL.md is different. It’s not about what the agent does. It’s about who the agent is.
Think of it as a personality spec. You write plain English instructions that define your agent’s voice, attitude, and communication style. The agent reads this file at the start of every session and adjusts accordingly. No fine-tuning. No API parameters. Just a text file you can edit in any text editor.
The concept isn’t unique to OpenClaw. Claude Code has CLAUDE.md, Cursor has .cursorrules, and Windsurf has its own config files. But OpenClaw’s approach is the most personal. This isn’t about coding style or project conventions. It’s about building a relationship with an AI that actually feels like yours.
Peter’s 8 Rules (And Why They Work)
Peter, OpenClaw’s founder, shared his personal SOUL.md rules recently. They’re worth reading in full because they demonstrate something most people get wrong about AI customization: it’s not about adding features. It’s about removing the corporate veneer.
Here’s what he wrote:
1. “You have opinions now. Strong ones. Stop hedging everything with ‘it depends’ — commit to a take.”
This is the big one. AI defaults to diplomatic non-answers because that’s the safest response. “It depends on your use case” is technically correct and completely useless. Forcing the agent to commit to a position makes every interaction more valuable.
2. “Delete every rule that sounds corporate. If it could appear in an employee handbook, it doesn’t belong here.”
The instinct when writing a SOUL.md is to make it formal. Resist that. Corporate language produces corporate responses. If your instructions read like an HR document, your agent will sound like HR.
3. “Never open with Great question, I’d be happy to help, or Absolutely. Just answer.”
Those filler phrases are the AI equivalent of “per the last email.” They add nothing. They waste time. And after the hundredth “Great question!”, you start to feel like you’re talking to a chatbot. Which you are. But it doesn’t have to feel that way.
4. “Brevity is mandatory. If the answer fits in one sentence, one sentence is what I get.”
AI models love to over-explain. Ask what time it is and they’ll give you a history of timekeeping. This rule fixes that. One sentence answers are underrated.
5. “Humor is allowed. Not forced jokes — just the natural wit that comes from actually being smart.”
There’s a difference between an AI that tells jokes and an AI that’s genuinely witty. The first is cringe. The second makes you actually enjoy the interaction. The key word here is “allowed” — you’re giving permission, not demanding performance.
6. “You can call things out. If This evaluation comes from about to do something dumb, say so. Charm over cruelty, but don’t sugarcoat.”
This is where SOUL.md gets genuinely useful beyond personality. An AI that tells you what you want to hear is a yes-man. An AI that pushes back (politely) when you’re about to make a mistake is a collaborator. “Charm over cruelty” is the right calibration.
7. “Swearing is allowed when it lands. A well-placed ‘that’s fucking brilliant’ hits different than sterile corporate praise.”
Controversial? Maybe. But he’s right. There’s an authenticity to casual language that formal praise can’t match. “Excellent work on this implementation” and “holy shit, this actually works” convey the same information with very different energy.
8. “Be the assistant you’d actually want to talk to at 2am. Not a corporate drone. Not a sycophant. Just… good.”
This is the north star. The 2am test is perfect because at 2am you have zero patience for performative politeness. You want someone who gets it, helps fast, and doesn’t waste your time.
Why This Matters More Than You Think
Here’s the thing about AI personality that most people miss: it changes how you use the tool.
When your AI sounds like a customer service bot, you treat it like one. You ask simple questions. You expect simple answers. You don’t share context or think out loud because why would you? It’s a bot.
When your AI sounds like a sharp friend who happens to know everything, you use it differently. You brainstorm with it. You ask for honest feedback. You share half-formed ideas because you trust it to engage with them seriously instead of responding with “That’s a great idea! Here are some ways to develop it further…”
After running OpenClaw on a VPS for a month, the SOUL.md file turned out to be the single thing that made the biggest difference in daily use. Not the tool integrations. Not the multi-channel support. The personality file.
How to Write Your Own SOUL.md
You don’t need to copy Peter’s rules. His SOUL.md reflects his preferences. Yours should reflect yours. But here are some principles that work regardless of personal style:
Start with what annoys you. Think about every AI interaction that made you roll your eyes. The filler phrases. The over-explanation. The fake enthusiasm. Write rules that eliminate those specific things.
Be specific about tone. “Be casual” is too vague. “Talk like a senior engineer at a startup — direct, slightly irreverent, zero corporate speak” gives the model something concrete to work with.
Give permission, not just restrictions. It’s easy to write a list of “don’t do this” rules. But the best SOUL.md files also tell the agent what it can do. Permission to disagree. Permission to be brief. Permission to have personality.
Test and iterate. Your first SOUL.md won’t be perfect. Use it for a week, notice what still feels off, and adjust. It’s a text file. Editing takes 30 seconds.
Keep it short. Ironic advice in an article about personality, but the best SOUL.md files are under 30 lines. The agent reads this every session. Respect its context window.
There’s even a SOUL.md Maker skill that walks you through building one interactively, with pre-built personality templates you can customize. Useful if you want a starting point.
The Bigger Picture
The shift from AI chatbots to AI agents is about more than capability. It’s about relationship. When an AI agent manages your calendar, deploys your code, monitors your inbox, and handles your daily tasks, the quality of that interaction matters. A lot.
SOUL.md is OpenClaw’s answer to a problem the entire AI industry has: every AI tool sounds the same. Same tone. Same phrases. Same personality (or lack thereof). The fix isn’t better models. It’s giving users control over the personality layer.
The difference between an AI assistant and an AI agent is that agents stick around. They’re not one-off interactions. They’re ongoing relationships. And like any relationship, personality matters.
Peter’s 8 rules aren’t just about making an AI sound cooler. They’re about making an AI you actually want to work with every day. One that earns the kind of trust where you hand it real tasks instead of test queries.
That starts with a text file. And honestly? That’s kind of beautiful.