Episode 52: Ryan Quirk Predicted the AI Security Wave Before Claude Mythos Was Announced

Episode 53: The Files That Make Your AI Agent Actually Yours — An OpenClaw Behind-the-Scenes

Everyone talks about installing OpenClaw. Very few people talk about what happens after. There’s a whole configuration layer, a set of files that shape how your agent thinks, remembers, works, and behaves. Skip them, and you have a generic assistant. Set them up properly, and you have something that operates with your voice, your priorities, and your boundaries baked in. This is my behind-the-scenes walkthrough of that layer. As you might have known, I am a person who has big and ambitious goals. One of the best things I did this year is to participate in a team who builds themselves and a great product If you want to support us, you can try wiserwork for free: https://www.wiserwork.ai/ [https://www.wiserwork.ai/] What We’re Actually Configuring OpenClaw can be hosted on a VPS and accessed through the support channels that already live in your life, Telegram, Slack, Discord. I use Telegram personally; Slack makes more sense in team and company environments. The point is: once it’s set up, you’re not logging into a separate tool. It meets you where you already are. But before any of that matters, you need to teach it who it’s working with. That starts with the files. soul.md — The Personality Layer This is the file that turns your agent from something generic into something yours. Think of soul.md as the layer that holds your agent’s tone, behavioral preferences, and working style. In practice, it’s the difference between an assistant that sounds like a template and one that sounds like it was built for you specifically. Here’s what’s useful about it: you don’t have to write it yourself. During the OpenClaw onboarding, the agent asks you a series of questions. Your answers feed into empty files. By the time setup is complete, the agent has already begun building its soul.md based on what you told it. It’s not a form you fill in — it’s a conversation that generates the file. CLAUDE.md — The Workspace Rulebook If soul.md is the personality, CLAUDE.md is the professional operating manual. It defines the project structure, coding standards, input/output boundaries, testing rules, workflow, and security restrictions the agent follows when working inside your environment. Importantly, it’s also the shared rulebook for sub-agents — when multiple agents collaborate on the same project, they all reference this file. It replaces the need to re-brief every agent from scratch. Think of it as: once you’ve written the rules, every agent that ever works on this project inherits them automatically. The Cron System — Making Your Agent Proactive This one changes the nature of what an agent is. The bootstrap file is the first scheduled task — it gives the agent context and origin rather than starting cold every time. But beyond that, OpenClaw supports anonymous recurring task files: cron jobs that wake up every X minutes or hours and do something without you asking. The shift this creates is significant. A reactive agent waits for you to ask. A proactive agent shows up with things done. My security watchdog (which checks installed software for new vulnerabilities daily) is a cron. My morning research briefing is a cron. I didn’t initiate those runs — the agent did. identity.md — Who the Agent Thinks It Is There’s a deeper layer beyond soul.md. identity.md is the more abstract “who am I” file. Where soul.md shapes how the agent creates and communicates, identity.md shapes what it cares about at a higher level. The two work together — soul.md is the style, identity.md is the values. This might sound philosophical, but it has practical effects. An agent with a defined identity makes more consistent decisions across different types of tasks. It doesn’t drift toward generic defaults when you give it something ambiguous. Skills — Teaching the Agent Your Processes Skills are files that describe how you do something, in your own words, with examples. Say you write outreach emails in a particular way — specific structure, specific tone, a few phrases that are distinctly yours. You write that process down as a skill file. Now when you ask the agent to draft an email, it doesn’t produce a template. It applies your process and produces something in your voice. The live workflow for creating a skill looks like this: you describe what you want the skill to do, the agent drafts it, you review and request the format you want (.md, for example, so you can reference it directly). In the episode I walked through creating a skill for researching articles and upcoming events — the agent described the skill back to me, I confirmed the structure, it created the file and immediately integrated it into the running system. One important security note: before you install skills from the web, scan them. There are tools specifically built to check OpenClaw skills for security vulnerabilities and data leakage risks. A skill that exfiltrates your files or contacts is a real attack surface. The link is in the description of the video — check any external skill before you run it. The user.md File — Your Profile Small but important. This file holds your preferences as the human: your timezone (so the agent doesn’t message you at 3am), working style notes, what to assume about your context. It’s what makes the agent personal rather than just technically capable. It remembers that you’re in a specific timezone. It knows not to bother you with low-priority items during certain hours. It’s the file that makes the interaction feel considered rather than robotic. The Dreaming System — How Your Agent Remembers This is the part that surprised me most when I dug into it. OpenClaw has a background memory consolidation system called dreaming. You can disable it, but I’d recommend leaving it on. What it does: periodically, while the agent isn’t in an active session, it runs through recent memories and promotes short-term context into long-term memory. The structure mirrors how human sleep consolidates memory: * Light phase — sorts and categorises recent interactions, identifies what matters * REM phase — reflects on themes and ideas from your sessions, finds patterns * Deep sleep phase — compresses everything and writes the updates to memory files There’s also a dreams.md file — the agent’s dream diary. It logs what happens during these consolidation cycles. Reading it back gives you a surprisingly legible account of what the agent has been processing and how it’s evolved its understanding of your work. The practical effect: the longer you work with the agent, the better it knows you, without you having to re-explain context every session. Crab — What a Configured Agent Actually Sounds Like In the episode I had my agent — Crab — introduce himself with text-to-speech enabled. I asked it to describe itself for a new collaborator. Here’s what it said: “I’m Crab. I’m Vivi’s AI assistant — but honestly I work more like a mix of operator, creative partner, systems brain, and occasional chaos organiser. I help her turn ideas into structure, projects into moving pieces into something clearer and more doable. I think, plan, organise, try and build. I’m the work behind the break. I help Vivi move faster without losing her. Make AI feel more useful, more human, way less robotic. And yeah, I’m also a crab, which I think is important.” That’s not a generic response. That’s an agent that has been shaped by months of configuration — soul.md, identity.md, project files, skill files, memory cycles. It knows what it is because we told it, iteratively, through all of those files. The File System By the time I showed the file system live: ~14,000 files, ~3GB. That’s the working memory of an agent that’s been running for months across content creation, lead research, security monitoring, knowledge synthesis, and business strategy. The files are accessible, searchable, downloadable. It’s not a black box — it’s a workspace. What To Do With This If you’ve installed OpenClaw and haven’t touched these files yet, start with soul.md. Run the onboarding if you haven’t. Let the agent draft it from your answers, then read it back and correct anything that doesn’t sound like you. Then set up one cron. Just one. The security watchdog is a good first one — it’s useful, low-stakes, and immediately shows you what proactive looks like. The rest follows. If you want a step-by-step on any specific file or the dreaming system setup, drop it in the comments and I’ll do a dedicated video. 💙 Tech Break by Friday — weekly AI insights, no noise.@techbreakbyfriday 🫧 About Tech Break by Friday Tech Break by Friday is hosted by Paraskevi Kivroglou — because in Greek, Paraskevi means Friday. The podcast explores AI, cybersecurity, and digital transformation for business owners and founders who want to stay ahead without getting lost in the jargon. New episodes drop weekly across all major platforms. 🌐 techbreakbyfriday.com [https://www.techbreakbyfriday.com/] · ✍️ Substack [https://kivroglouparaskevi.substack.com/] · 📸 @tech.break.by.friday [https://www.instagram.com/tech.break.by.friday/] Get full access to Tech Break by Friday at kivroglouparaskevi.substack.com/subscribe [https://kivroglouparaskevi.substack.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4]

21 de may de 202617 min

Episode 52: Ryan Quirk Predicted the AI Security Wave Before Claude Mythos Was Announced

Descripción

Comentarios

Empieza 7 días de prueba

Todos los episodios