I’m writing this with voice typing. Dictating to a machine so I can talk to you about machines.
If that breaks your brain a little, you’re in the right place.
The reason isn’t laziness, though I have that in abundance. It’s because the whole point of this blog is figuring out where these tools actually deliver and where they’re expensive noise. The only honest way to find out is to use them. Document as you go. Report what’s real.
What I Actually Built
A few weeks back I put together some custom agents that compressed roughly three days of manual work into four hours. Not a demo. Not a proof of concept with curated inputs. Actual work, actual output, actually shipped. That got my attention in a way vendor pitches haven’t in a long time.
So I kept pulling the thread.
The blog runs on automation now. Gemini CLI handles publishing. Google Sheets manages slugs and headlines, linked to Google Docs for drafts, with a nightly job that publishes at 2am. It’s not pretty. It runs without me touching it.
The Lab
I run a virtualization setup in my home office. Multiple Docker containers, a few pods. Nothing that impresses anyone at a conference, but enough to actually build and break things on my own time. The goal is to wire in my Anthropic and Google subscriptions as the reasoning layer and automate workflows that actually need automating.
Not demos. Real workflows. Things that take time and shouldn’t.
I’ve been in IT long enough to watch every hype cycle build and crater. Cloud was going to eliminate the sysadmin. Kubernetes was going to simplify deployments. Zero Trust was going to solve security. All of these things have real value buried under the marketing. None of them delivered on the pitch.
AI feels different to me. Not because the vendors got more honest — they didn’t. But because I’m actually compressing real work with it, and the gap between what it promises and what it delivers is smaller than anything I’ve seen in thirty years of doing this. That’s not nothing.
Gemini vs Claude: Honest Take
Running both day to day. Here’s what I’ve found:
Gemini is genuinely good at research. Pulls from the web, synthesizes sources, finds specific details buried in documentation. For “what’s the current state of X” questions, it’s the right tool. It actually uses the internet rather than pretending to know things from a training cutoff.
Claude is better at code. Meaningfully better. It hits one-shot deployment confidence more often. Gemini can write code, but it needs more steering. For a codebase that actually has to run, Claude gets there faster.
Strategy: Gemini for research, Claude for building. Route the work to the right tool. Don’t use the expensive model for everything just because it’s there.
What’s Next
MCP servers on Docker. How to build them, host them, and keep the token costs from running wild while using both providers. And security. If you’re wiring an AI agent into your systems and handing it tool access, you need to think hard about what you’re actually doing. Most people don’t. I’ve watched that movie enough times to know how it ends. More on that soon.
