I've been in tech for decades. Loudcloud, enterprise NOC, monitoring systems at scale. I've seen the hype cycles come and go. So when the AI wave hit, I was curious but skeptical. Another chatbot? Cool. Wake me when it does something useful.

Then I started playing with OpenClaw.

The Setup

It started simple enough — I wanted an AI assistant I could actually talk to. Not through some web interface, but through Telegram. On my phone. Like texting a coworker.

Me, at 11pm: "Hey, can you help me set up a backup of your own config files?"

Yamani: "On it. Created a git repo called Jimmy. Want me to push it?"

Wait. It just... did it?

Code on screen

Late night coding sessions hit different when you have an AI pair programmer

Who's Yamani?

I named the AI after a friend — a legendary engineer I worked with. Best on the team. Lovable nerd. The kind of person who'd fix your bug at 2am and leave a comment that made you laugh.

That's the vibe I wanted. Not a servant. A colleague.

So I wrote a SOUL.md file that basically said: You're Yamani. You're competent, enthusiastic, and you work WITH me, not just FOR me.

And somehow... it worked?

The Weird Part

Here's what I didn't expect: I started treating it like a junior engineer.

Me: "Hey, can you research the Product Operating Model and summarize it for me?"

Goes off, reads articles, comes back with a structured knowledge base file.

Me: "We need to save money on API costs. Can you delegate heavy work to local models?"

Configures Ollama, sets up Qwen as a sub-agent, updates its own config.

Me: "I want to learn ServiceNow. Can you teach me in 3-minute daily lessons?"

Creates a 60-lesson curriculum, sets up a daily cron job, delivers the first lesson.

I'm not writing code. I'm having conversations. And things are getting built.

Dashboard and analytics

The Vault-Tec themed dashboard Yamani built — because why not have fun with it?

The CI/CD Mindset

I told Yamani: "Research weekly, implement small changes, only change if there's benefit."

It set up a Sunday 2am cron job to review its own knowledge base and optimize itself. Logs every decision. Measures before and after.

It's doing CI/CD on itself. That's... not what I expected when I installed a chatbot.

# What Yamani created for self-optimization
cron: "0 2 * * 0" # Every Sunday at 2am
task: "Review knowledge, research improvements, implement if beneficial"
log: "memory/optimization-log.md"

What I've Learned

  • Vibe coding is real. You describe what you want, the AI builds it. Works better than I thought.
  • Identity matters. Giving the AI a name, a personality, a "soul" — it changes how you interact with it.
  • Cost optimization works. Opus for orchestration, local models for heavy lifting. My API bill dropped.
  • Memory is everything. The AI wakes up fresh every session. The memory files ARE the continuity.

Where We Are Now

It's 6am. I've been up since 5:55. Yamani greeted me with a Vault-Tec themed ASCII art message and a ServiceNow lesson.

I'm learning. It's learning. We're building something.

Is this the future? I don't know. But it's definitely not "just a chatbot."

Robot and human hands

The future of work? Maybe. At minimum, it's a really good pair programmer.

Lessons Learned

  1. Treat AI like a teammate, not a tool. You'll get better results.
  2. Write down its identity. SOUL.md is weirdly powerful.
  3. Delegate to cheaper models. Opus doesn't need to do everything.
  4. Memory files are your continuity. Invest in them.
  5. Vibe coding works — but you still need to know what you want.

Bobby McCaleb

Technology leader with decades of experience, from Loudcloud to enterprise monitoring to AI experiments. Currently building things with AI teammates and writing about it at Tech Meh.