Back From Pause: What Two Weeks Away Taught Me About Building With AI

It feels incredibly good to be back.

Two weeks off.
One week skiing.
One week actual vacation.

And now — finally — building again.

There’s something powerful about stepping away completely. When you return, the noise is gone. What remains is clarity.

What actually moves things forward?
And what was just technical curiosity?

The Ubuntu Server Experiment

During the break, I built something I had wanted to test for a long time:

A physical Ubuntu server running OpenClaw.

It was a proper build.

Fresh Linux install.
SSH keys.
Firewall.
Docker.
Token isolation.
Local-only binding.

It worked.

And it was genuinely educational.

But here’s the honest conclusion:

It does not help me solve the problems I’m currently working on.

It was a valuable side quest — not an accelerator.

Right now, vanilla coding with Claude Code is simply more effective for my workflow.

That realization matters.

A Surprising Tooling Shift

One of the clearest takeaways from the OpenClaw project was this:

Claude Code is significantly stronger at terminal-level reasoning than ChatGPT.

It’s noticeable in real work.

Especially when:

  • navigating environments
  • debugging infra
  • working inside Docker
  • reasoning about system state

More and more, I find Claude replacing my ChatGPT use cases in development workflows.

Not emotionally.
Not ideologically.

Just pragmatically.

I use what works best.

Turning Everything Back On

Today I restarted the actual projects.

Supabase had paused all databases due to inactivity.
They’re live again.

And once again, I’m impressed by how smooth the Vercel + Supabase stack feels.

Deploy.
Database.
Auth.
Edge functions.

Minimal friction.

That’s the kind of stack I want:
Infrastructure that disappears.

Building Layer Two of the AI Council

Today we built the second layer of our AI Council.

We tested the full creation flow end-to-end.
And we added image support.

For now, there’s no internal model competition on images.

Nano Banana handles rendering until other platforms release stronger image models.

The workflow is simple:

Claude Code prompts.
Nano Banana draws.

And it works.

There’s something deeply satisfying about seeing systems connect.

Not ideas.
Not plans.
Working chains.

The Real Progress

Progress isn’t in the vision.
It’s not in the architecture diagrams.
It’s not in the server builds.

It’s in iteration.

In turning things back on.
In running the flow again.
In shipping the next layer.

It’s very good to be back.

Now we keep building.