I sat down with Brian Scanlan to explore what “AI-first” development actually looks like inside Intercom’s 15-year-old Rails monolith.
They’re generating over 95% of their code with Claude Code… and it’s not just engineers. Teams across the company are involved.
We got into how this is reshaping code review, who’s running production queries, and how they’re building guardrails to keep things from drifting.
The part that stuck with me… knowing when to disengage the autopilot.
I'm hearing the same conversation play out at small software companies right now. Engineers telling their managers: am I being forced to adopt this, or do I need to go somewhere else?
Wrote about why I think the job description already changed, and what I'd rather see teams do about it than have that exhausting conversation on repeat.
I know this because I'm the guy in the middle. Let me explain.
2006: Trey Parker and Matt Stone write "Oh My Science!" into two Season 10 episodes... a joke about replacing religious expressions with scientific ones.
2008: A coworker and I think this is hilarious. We build a Twitter side project called Oh My Science. It fetched tweets from the API and awarded gold stars to anyone caught thanking science for something they might have otherwise credited to a deity.
2009: I have a messy folder of Zsh configuration files. I want to convince a few coworkers to install it on their laptops so I don't have to remember all the verbose git commands we kept typing while pairing on their machines. I need a name for the repo. I glance over at our oh-my-science private repo. I shrug. "oh-my-zsh" it is.
Today, there are 8,400+ public "oh-my-..." repositories on u/GitHub.
There's a decent chance you've used one. There's a chance you've built one.
If you have, I'd love to know... did you know where the name came from?
That’s really funny. I think I’ve seen one of these repos before, but I don’t think I’ve used one and I haven’t created one. Not too late to hop on the bandwagon I suppose!
Built a RAG tool in Ruby to surface historical context from Jira, Confluence, and GitHub before starting work on client tickets. No prior experience with embeddings or vector databases. Ended up adding an MCP server so it works directly inside Claude Code.
I wrote this after noticing how much framework discussion focuses on greenfield work. In practice, most teams I see are inside 10 or 12 year old systems, evolving them under real constraints.
The piece is about that “second act” of software. After launch. After early growth. When reliability and discipline matter more than novelty.
