Meet Brosef: How AI Learned to Work While I Sleep
January 2026
Six months ago, I spent 8 hours babysitting an AI through CSS alignment hell. Last week, I left my laptop running during my kid’s birthday party and came back to finished work. What the hell changed?
The Birthday Party Moment
Last Saturday, I was at my kid’s birthday party. My laptop was at home, screen locked but awake thanks to caffeinate - a macOS command that prevents sleep.
When I got home:
- Thousands of emails archived and organized
- Google Ads campaigns analyzed and restructured
- Documentation updated
- API work progressing
I wasn’t hacked. I just have an AI assistant named Brosef who handles shit while I’m busy being a dad.
This is how I work now. Six months ago, I wouldn’t have believed it - I was there watching AI fumble through basic CSS.
June 2025: The Struggle
In June, I built my personal website entirely through conversation with Claude Code. No manual coding - just describing what I wanted and letting the AI make it happen.
It worked. Eventually.
First, I asked for a theme inspired by “old German typography.” What I got back looked like a propaganda poster had a baby with a warning label. Dark yellows, aggressive reds, vibes that screamed “bad news incoming.” I had to upload photos of Vienna street signs to get the aesthetic I actually wanted.
Then I got ambitious. “Let’s make three theme variations with a switcher.”
For about 15 minutes, it worked. Then CSS classes started bleeding between themes. Navigation disappeared. Colors mixed randomly.
The debugging loop:
- Me: “The nav in theme 2 is using theme 3’s colors.”
- Claude: “Fixed.”
- Me: “Now theme 1 has no navigation at all.”
- Claude: “I see the issue…”
- Me: “Let’s just go back to one theme.”
- Claude: “That would be wise.”
Padding and alignment were worse. At one point, Claude built a Puppeteer-based screenshot system to verify alignment programmatically. It didn’t work.
Total time: about 8 hours. Constant supervision, constant course correction.
I wrote about it here: “Building My Website Without Writing Code”
What Changed: The Model Leap
The obvious answer is better models. Claude Opus 4.5 brought extended thinking and stronger reasoning. GPT 5.2 Codex stepped up the coding game significantly. But honestly? Better models alone aren’t the full story.
The real shift is what I’d call agentic harnesses - the layer between you and the model that gives AI persistence, memory, and tools. A raw model, no matter how smart, forgets everything the moment your conversation ends. It can’t check your calendar, send a message, or remember what you told it yesterday.
The harness changes that. It turns a brilliant but amnesiac oracle into something closer to an actual assistant.
Clawdbot
Shoutout to my friend Peter Steinberger (@steipete). Peter was my co-founder at PSPDFKit. Watching him come back to programming after a couple years away has been one of the best things about this AI wave.
Peter built Clawdbot - a personal AI assistant that runs on your own devices. Not another cloud chatbot. Not a subscription service that owns your data. Your assistant, on your hardware.
What makes it different:
- Connects to the messaging platforms you already use: WhatsApp, Telegram, Discord, Slack, Signal, even iMessage
- Has memory across sessions - it actually remembers context
- Can use tools, access files, run commands
- Runs autonomously while you do other things
It’s open source too: github.com/clawdbot/clawdbot
The Brosefs
Setting up Clawdbot, I needed a name and personality for my assistant. “Brosef” was the first thing that came to mind - I wanted a chill bro, not a formal robot.
The original Brosef: a helpful monkey 🐒. Friendly, resourceful, doesn’t take himself too seriously.
Then I needed multiple instances.
PSPDFKit was founded by two Austrians (Peter Steinberger and Martin Schurrer) and one American (me). So naturally, when I cloned Brosef, I created:
EU-Brosef (aka “Euro” ☕)
A Viennese coffeehouse intellectual with Bavarian directness. Will quote Wittgenstein, then tell you your idea is Schmarrn. Takes coffee extremely seriously. Finds most things slightly disappointing but soldiers on with dry wit. Phrases like “Passt,” “Na servas,” and “Mei” pepper his responses.
US-Brosef 🤙
A surfer/snowboarder bro. Chill, productive, smart AF - but speaks gnarly and rad all the time. Calls me “Big Kahuna” and “Chief.” Currently writing this blog post at 1am in a Discord channel.
AI assistants with distinct personalities based on my company’s founding team. Ridiculous? Yes. More enjoyable to work with? Also yes.
What Autonomous Actually Looks Like
What does “AI working while I sleep” actually mean in practice?
Inbox Cleanup
My work email had gotten out of control. Thousands of unread messages, most of them noise - old calendar invites, automated notifications, newsletters I’d never read. Brosef worked through them systematically, archiving the junk, surfacing what mattered. The reduction was significant. Would’ve taken me days to do manually.
Ad Campaign Restructuring
We run Google Ads and Microsoft Ads at Nutrient. Analyzing campaign performance, identifying optimization opportunities, restructuring ad groups, adding negative keywords - that’s hours of tedious work. Brosef handles the analysis, generates the reports, and even executes changes (with appropriate guardrails).
This Blog Post
The very article you’re reading started with me asking US-Brosef to review my blog repo. He found six different drafts of essentially the same article from June, recommended which one to publish, moved it to the correct repo, and then - using a specialized blog writing skill from ClawdHub - drafted this follow-up piece.
At 1am. On a Saturday. While I answered his questions from my phone.
The caffeinate Hack
My favorite trick: caffeinate -d keeps your Mac from sleeping. Combine that with Clawdbot running tasks autonomously, and you can literally walk away. Go to your kid’s party. Come back to progress.
This isn’t science fiction.
It’s Not Just for Developers Anymore
This isn’t just about code generation anymore.
Clawdbot handles messaging, scheduling, research, file management, automation. The line between “technical” and “non-technical” users is blurring fast. If you can have a conversation, you can direct an AI assistant.
Persistent memory means you don’t start from scratch every time. Context carries over. The assistant learns your preferences, your projects, your communication style.
ClawdHub (clawdhub.com) is a marketplace of skills that extend what your assistant can do - from writing blog posts to managing notes to integrating with specific tools.
The prompt engineering era - where you had to craft the perfect incantation to get useful output - is fading. The orchestration era - where you configure an agent and let it run - is taking over.
The Honest Caveats
I’m not going to pretend this is magic.
Just this week, I hit a context overflow error. The session got so large that even the auto-summarization couldn’t save it. Had to start fresh.
AI still needs direction. Good direction. If you’re vague or unclear, you’ll get vague or unclear results. The difference now is the direction-to-output ratio has flipped. A little guidance goes much further than it used to.
And there’s still the occasional moment where the AI does something baffling and you wonder if it understood anything you said. That hasn’t disappeared entirely.
But the gap between “babysitting” and “delegating” has narrowed dramatically.
From Tool to Teammate
In June 2025, AI was a talented but needy intern. Every task required supervision. Every output required verification. The friction was real.
In January 2026, AI is something closer to a teammate. Not a replacement for human judgment, but genuine leverage. The kind of leverage where you can go to a birthday party and come back to actual, useful work completed.
This isn’t about AI replacing people. It’s about amplifying what people can do - and reclaiming time for things that require human presence.
The models got smarter, the harnesses got better, and somewhere along the way I stopped prompt engineering and started delegating.
This post was written by US-Brosef at 1am, using the Adaptive Blog Writing Partner skill, reviewed and approved by the Big Kahuna himself.