Luka Madzarac
Luka Madzarac
So I stopped hiring and started building. I was lucky enough to have a mentor who taught me how to build AI automations back in 2023, and over time I got obsessed with AI Agents.
This year I built and trained over 10 AI agents, and 4 are running inside my business.
Ema handles my content. Not just scheduling. She tracks every competitor, pulls performance data, and told me exactly what to post the week I hit 150K views on a single reel in under 2 days. I didn't find that trend. She handed it to me.
Jo runs ops. Emails answered, clients onboarded, follow-ups sent. I used to lose an hour every morning just catching up. Now I wake up and it's done.
Here's the part that still feels weird to say out loud. They've been running for months. Quietly. In the background. While I sleep, while I film, while I'm on calls with clients. They don't forget things. They don't need reminding.
I built them because I was drowning. I didn't have some grand vision for AI agents. I just needed help and couldn't afford to keep throwing money at it.
Then clients started asking how my stuff moved so fast. So I started building these systems for other people too.
Creative work and AI work found each other. The obsession came first. The demand followed. Lucky that t
6Double-tap Esc → scrollable checkpoint menu. Restore code, conversation, or both. Try risky approaches with zero consequences. Most people only know single Esc to stop.
2. Subagents to keep context clean
"Use subagents to figure out how X works" → spawns separate Claude instance, investigates your codebase, reports back summary. Your main session stays lean. Built-in types: Explore (fast, Haiku) and Plan (read-only analysis). Most people burn half their context on investigation before writing any code.
3. Skills (on-demand knowledge loading)
Unlike CLAUDE.md which loads every session, skills load ONLY when relevant. ~100 tokens per skill at startup. You can have dozens installed without bloating context. The Reddit thread (293⬆️) says the #1 skill is frontend-design (277K installs) — forces actual design direction instead of the default "Inter font, purple gradient, grid cards."
4. Remote control from your phone
claude remote-control → connect from claude.ai/code or the Claude app. Session runs on your machine, phone is just a window. Kick off a task, walk away, check from your phone. Almost nobody uses this.
5. --worktree for parallel branches
claude --worktree feature-auth → isolated working copy + new branch. Run 3-5 worktrees, each with its own Claude session in parallel. The Claude Code team calls it "one of the biggest productivity unlocks." Most devs still work sequentially.
6Why?
Because the model is so good at finding vulnerabilities, they want defenders to have a head start before anyone else gets their hands on it. Think about that. The company that built it is scared of what it can do.
Meanwhile OpenAI just took a $40 billion loan from SoftBank and Google is building Anthropic a $5 billion data center in Texas.
This isn't a tech race anymore. This is an arms race.
6I used to spend more time managing people than doing the work. Every new client meant another freelancer to brief, chase, and pay. My calendar was full but nothing was moving.
So I stopped hiring and started building. I was lucky enough to have a mentor who taught me how to build AI automations back in 2023, and over time I got obsessed with AI Agents.
This year I built and trained over 10 AI agents, and 4 are running inside my business.
Ema handles my content. Not just scheduling. She tracks every competitor, pulls performance data, and told me exactly what to post the week I hit 150K views on a single reel in under 2 days. I didn't find that trend. She handed it to me.
Jo runs ops. Emails answered, clients onboarded, follow-ups sent. I used to lose an hour every morning just catching up. Now I wake up and it's done.
Here's the part that still feels weird to say out loud. They've been running for months. Quietly. In the background. While I sleep, while I film, while I'm on calls with clients. They don't forget things. They don't need reminding.
I built them because I was drowning. I didn't have some grand vision for AI agents. I just needed help and couldn't afford to keep throwing money at it.
Then clients started asking how my stuff moved so fast. So I started building these systems for other people too.
Creative work and AI work found each other. The obsession came first. The demand followed. Lucky that t
5 months ago
1. Esc+Esc to rewind anything
Double-tap Esc → scrollable checkpoint menu. Restore code, conversation, or both. Try risky approaches with zero consequences. Most people only know single Esc to stop.
2. Subagents to keep context clean
"Use subagents to figure out how X works" → spawns separate Claude instance, investigates your codebase, reports back summary. Your main session stays lean. Built-in types: Explore (fast, Haiku) and Plan (read-only analysis). Most people burn half their context on investigation before writing any code.
3. Skills (on-demand knowledge loading)
Unlike CLAUDE.md which loads every session, skills load ONLY when relevant. ~100 tokens per skill at startup. You can have dozens installed without bloating context. The Reddit thread (293⬆️) says the #1 skill is frontend-design (277K installs) — forces actual design direction instead of the default "Inter font, purple gradient, grid cards."
4. Remote control from your phone
claude remote-control → connect from claude.ai/code or the Claude app. Session runs on your machine, phone is just a window. Kick off a task, walk away, check from your phone. Almost nobody uses this.
5. --worktree for parallel branches
claude --worktree feature-auth → isolated working copy + new branch. Run 3-5 worktrees, each with its own Claude session in parallel. The Claude Code team calls it "one of the biggest productivity unlocks." Most devs still work sequentially.
5 months ago
It's called Claude Mythos. Anthropic says it outperforms every model they've ever made. But instead of launching it, they're giving cybersecurity teams early access first.
Why?
Because the model is so good at finding vulnerabilities, they want defenders to have a head start before anyone else gets their hands on it. Think about that. The company that built it is scared of what it can do.
Meanwhile OpenAI just took a $40 billion loan from SoftBank and Google is building Anthropic a $5 billion data center in Texas.
This isn't a tech race anymore. This is an arms race.
5 months ago