Yeah, that title is clickbait. But I’m only like 10% sorry because it’s mostly true.

Yes, AI and Vibe coding is messy. Ship it anyway (once you lock it down).

I’ve been writing code for over 30 years. Lately I’ve been building with agents, and the code they produce can be insecure, ugly, hard to maintain, inconsistent. The patterns are weird. The variable names are questionable. The architecture sometimes feels like it was designed by committee during an earthquake. I’m not going to argue with any of that. It just doesn’t matter as much as you want it to if you address it.

How I use GitHub Copilot and Obsidian as executive function scaffolding for my AuDHD brain.

Resources

• Copilot CLI
• Obsidian
• Plugins: Tasks, Dataview, Templater, Daily Notes (core)

The Problem Nobody Talks About

I’m an Engineering Manager. I have direct reports, a dozen active projects, weekly 1:1s, sprint planning, cross-team coordination, incident response, weekly updates and the kind of context-switching that would make Chrome crash.

I wanted a blog. I didn’t want to manage one.

The requirements: write markdown in Obsidian, tell my AI assistant to publish it, have it show up on my domain. No local Hugo install. No Ruby. No Node build step. No “oh wait, I’m on my other laptop and don’t have the toolchain.” Just markdown and a conversation.

The workflow

1. Write in Obsidian (where I already live for notes, planning, everything)
2. Tell Copilot CLI to publish it
3. Copilot copies the markdown to the blog repo, sets up frontmatter, commits, pushes
4. GitHub Actions builds the Hugo site and deploys to Pages

That’s it. I never run Hugo. I never think about the build. I write, I say “publish,” it’s live in 30 seconds.

I’ve kept my dotfiles synced across machines for years. Shell configs, git aliases, editor settings; the usual. But when I looked at my setup recently, I realized I’d completely ignored the newest layer: all the AI CLI tools I now rely on daily.

GitHub Copilot CLI configs, Claude Code settings, custom instructions, memory files, tool preferences. None of that was being tracked. Every new machine meant rebuilding that context from scratch.