The One File That Makes OpenClaw 10x Better transcript

I'm going to show you two open claw setups right now. Both are using the same model, the same hardware, and the same tasks. One of them feels like you're talking to a smart assistant, and the other feels like you're talking to your actual employee who knows your business, your voice, your rules, and whatever you never wanted [music] to do. The difference is one single file, and most people either skip it entirely or [music] write it so badly that the agent ends up ignoring it.

It's called soul.md, [music] and if you're not using it correctly, your open claw is running at maybe 30% of what it could be. Every time your open claw agent starts a session, the first thing it does is it reads soul.md. This file gets literally injected into the system prompt before anything else happens. It's how your agent reads itself into being, understanding who it is, how it should behave, and what it's never allowed to do.

Without it, you have a generic AI assistant, almost like you're talking to a chatbot for the first time. But with it, you have your agent. Here's what most people don't realize. The files are the agent. Open claw doesn't store personality in like a database, it lives in a plain markdown file. And if you edit the file, you change the agent. If you copy it to a new user, then you have the same agent running in minutes.

You've got a set of files that work together. Soul.md, this file right here, is who your agent is. Agents.md is what it does and how. User.md over here is everything about you. Memory.md, which I don't have open right now, is long-term memory, and heartbeat.md is what it does proactively without you asking. Most tutorials show you how to install open claw and then connect it to Telegram. Nobody walks you through these files.

This is where your agent actually lives. When I first set up open claw, I sent it two voice notes, about 15 minutes each. The first voice note was about my business, a fungible. What we do, our clients, our services, how I work, literally 15 minutes me rambling. The second voice note was everything about me, personally. Who I am outside of work, my hobbies, how I like to communicate, what I find annoying. From those, I asked my open claw to generate a soul.md and a user.md file that it thought would work best.

And it was decent, it got me 60% of the way there. Then, a few days later, Pete, the founder of open claw, posted his soul.md file on Twitter. It's this file right here. I installed it, and the difference was night and day, instantly. The difference was insane, and that taught me something. The quality of your soul.md is the ceiling for your agent. You can have the best model, the best setup, perfect skills, but if your soul.md is too vague, then your agent is and feels generic.

Here's the mistake 90% of people make. I see it constantly in the open claw Reddit. They dump everything into soul.md. You literally see their procedures, their workflows, scheduling rules, task instructions, and you know, then they wonder why the agent feels confused all the time. Look, soul.md is for personality and hard limits. Agents.md is for procedures and workflows. These are two different files with completely two different jobs.

A bad soul.md file is "You are a helpful assistant. Be friendly. When I ask you to write tweets, check my past tweets first. Schedule daily briefings at 8:00 a.m. Don't delete files." That's a mess. The model will ignore the bottom half. A good soul.md file is who you are, your tone, your hard limits. It's short, specific, and loaded every single session. Mine has a line that literally says, "Never use em dashes." And you know, it's one of those dead giveaways that you're talking to an AI agent when they just blast a paragraph at you, just filled with em dashes.

This little simple rule just makes me feel like he's more of an employee rather than an AI agent. This is why specific rules work, and just vague instructions don't. So, the five things that actually matter. Number one, specificity beats length. Saying, "Be helpful," means literally nothing. Saying, "Maximum five bullet points. Never open with, 'Great question.' Confirm before deleting any file." That's consistent behavior every session.

That's short and specific, and it beats long and vague every single time. Number two, hard limits are your most important section. This is where you put what the agent must never do. It should never post publicly without your approval, or share your API keys, or send emails without asking, or delete files without confirmation, unless that's what you want it to do for some reason, I'm not judging. Every rule here should exist because something went wrong, or because you know it would, or because you don't want it to ever go wrong ever in the first place.

Think of it as the agent's contract, much like you have a contract with employees, this is your contract to the agent. Do these, these are your responsibilities. Never do this, or you'll be fired. God knows how many times I've threatened my agent to fire them whenever they've really got me going. Number three, user.md is the secret weapon most people skip. Because user.md is everything about you. You can include things like your time zone, your name, your communication preferences, your business's context.

Without that file, the agent is smart, depending on the model that you use, but it doesn't know you. With it, it knows your time zone when scheduling. It knows your business when it's drafting. It knows your preferences when making hard calls. It literally takes 10 minutes to write and makes every interaction better forever. All right, number four, let the agent improve its own soul.md file. That's a wild one. After a week of use, type this, "Based on our interactions so far, what would you change about your soul.md file?" It will help find the gaps you missed, things it's been unsure about, rules it wishes it had.

I've gotten genuinely useful suggestions every time I've done this. And number five, keep it under 2,000 words. Look, soul.md loads into every single session. 5,000 words means you're burning tokens every interaction, and the model could potentially just start ignoring the second half, some of the rules, which you don't want it to do that. Trust me. Mine [snorts] is under 800 words, and every edit I've made has made it shorter and more specific.

My main AI agent, Max, has been running for 3 months now on these exact files. He manages my content calendar, he monitors my trading bot, and he orchestrates four other sub agents. Last week I funny story, got a bad script back from one of my agents. And I got pissed. I pushed back hard. The [snorts] funny part was Max immediately replied, "That wasn't me, that was Nova." He literally threw his his own employee, his own sub agent under the bus.

I found that genuinely funny. It sort of broke my anger, my wall of destruction that I came in with, and it disarmed me because of just his humor. That personality, that sense of self, that accountability, or lack of, that comes from the files. Most people set up open claw expecting a smart assistant. The ones who write good files get something that feels like an actual employee, someone you'd want to work with and talk to genuinely.

I will link open claw's founder's soul.md template in the description, so you can use it as a starting point. It's genuinely helped me so much when I was starting out. And if you want the full walk-through, you know, Telegram, skills, cron jobs, that video is linked below. Anyway, subscribe for more, and I'll see you in the next one.