As of October 2026, I run 19 apps. Every month I ship more than six improvement releases and launch four new apps.
When one person runs that many apps, doing the work by hand simply doesn't scale. I have to commit fully to making decisions.
The term "skill ops" came up in a conversation with an AI. A Google search turns up almost nothing, so it seems to be a word that happened to be coined on the spot. I liked it a lot.
Skill ops.
It means running operations on the assumption that skills do the work. At a previous job, for example, people put real effort into building a data platform, and humans collected and analyzed the data. If I tried to do that alone, operations would grind to a halt.
I just set up the Firebase and Google Analytics foundation and leave the analysis to AI.
- Data analysis takes one line: "Analyze this."
- Log analysis takes one line too: "Check this."
The human's job is to look at the insights that come out, decide what to do, and give the instructions.
I run sites called Tenkaichi and Buzz Spot. They are early in their life, so naturally there isn't much content yet. If nobody posts, I have to post myself until the sites fill up.
Still, if I wrote every article myself, operations wouldn't keep up.
- "Research this restaurant and put together a post about it."
- "Who's stronger, Yamcha or Ryomen Sukuna?"
I set things up so that asking the AI a question like these takes it all the way to a published post, automatically. SEO is covered too. Updating sitemap.xml and requesting indexing in Search Console both run on skill ops.
What I've learned from skill ops carries over to enterprise development as well. Until now, when an incident happened, we investigated it like this:
- Check the logs in CloudWatch or Elasticsearch.
- Track down the code that produced the log.
- If it's outside my team's area, ask the owning team on Slack to investigate.
- Fix the bug through trial and error in a local environment.
Instead, keep the backend, frontend, and infra repositories on your machine, and let the AI handle everything from reading the logs to digging through the repositories. Of course, letting Claude Code read all of your production logs is a security problem, so you need to limit it to the development environment or hand it excerpts. Even so, operations built around AI agents work perfectly well in the enterprise.
Or rather, I expect that from now on, unless we design all of our work around AI agents, we'll start to feel the pain in productivity. Humans can do it if they push hard, but it will feel like a chore. That's because we now know how convenient AI is.
To let humans focus on decisions, it pays to design operations around AI from the start.