Time for an honest meta-moment: this blog about AI was largely built with AI. Specifically with Claude Code — Anthropic's AI coding assistant that runs directly in the terminal. Sounds like self-congratulation. It isn't, I promise.
What is Claude Code anyway?
Claude Code isn't a chat interface — it's an AI assistant that works directly in the terminal. It reads files, writes code, runs commands, and does all of this on your actual system, not in a sandbox. That makes it powerful. And yes, it requires a bit of trust.
You give it a task, it works through it independently, asks when something's unclear, and proposes changes before making them. In practice it feels like a very fast, always-available developer — who occasionally makes peculiar decisions.
What AI did here
The technical setup: A custom PHP static site builder — no off-the-shelf framework, built to measure. Folder structure, multilingual support, sitemap, RSS, template system — Claude Code built all of that while I was drinking coffee.
The pipeline: GitLab CI with automated deployment, image generation via Stable Diffusion, LLM code review — Claude Code sketched it out, I reviewed and approved (or corrected).
Article drafts: Yes, the articles come with AI assistance too. I provide the topic, tone, and key points — Claude writes a draft. Then I read, revise, cut, add. What you're reading isn't 1:1 AI output, but it's also not written entirely by hand.
What the human did
And this is the important part:
Decisions: What is this blog? Which topics? What tone? Who's the audience? These aren't technical questions — they're editorial decisions. I make those.
Control and approval: Claude Code does nothing I don't sign off on. Every article, every configuration change — I read it before it goes live. Sometimes I change a lot. Sometimes a little.
The perspective: The fact that I use OpenRouter on a paid tier because I care about privacy. That I run Ollama myself. That I use ChatGPT daily but don't trust it blindly. That comes from me, not from an AI that doesn't know me.
What went wrong
The unvarnished truth: over 60 merge requests and more than 500 pipeline runs before the first post went live fully automatically without errors.
That sounds like chaos. It was... iterative.
Missing packages in the Docker container. Different ImageMagick versions between local build and CI. Misconfigured volume mounts on the GitLab runner — paths that apply on the host, not inside the container. A shell runner suddenly acting as a Docker runner. SSH keys that need to be in the right section of the config, not just the right file.
Claude Code analysed every one of these errors and suggested fixes — which usually worked. And sometimes uncovered the next problem.
On top of that: an article draft that came out too formal, a config that didn't build on the first try. Small stuff compared to the pipeline odyssey. But honesty demands mentioning it.
What I take away from this
AI as a coding assistant isn't hype. It's real productivity gain — especially for individuals without a development team behind them. I could have built this blog alone, but it would have taken longer and been worse in places.
500 pipelines sounds like a lot. But it's also 500 rounds of immediate feedback instead of weeks flying blind. That's the real advantage: not fewer mistakes, but faster learning from them.
AI doesn't replace the person steering it. It accelerates execution. That difference sounds small. It isn't.
A blog about AI that speaks honestly about AI — built with AI, across 60+ merge requests and 500+ pipelines. That feels oddly fitting.
And if you're wondering whether to try AI tools for your own projects: do it. Getting started is easier than it looks — and you learn the most by simply beginning.