I guess I'm AI-pilled now?

As I’ve mentioned before, I have pretty mixed feelings on AI. Like many things in life, more than one thing can be true.

I’m also going to throw a caveat here up front that most of my examples use GitHub Copilot. I work for GitHub so I think this makes sense. I’m not trying to sell the tool to you—feel free to pick whichever tool you want. This is just the one that is cheapest for me to use!

I spent the recent long weekend binge-listening to the How I AI podcast. I thought before that weekend that I was pretty in-the-know on AI tools at this point, but gosh was I wrong. I listened to several practical ways folks are using AI and it really opened my eyes to just how many options there are. I guess that’s why folks refer to the advantage for younger folks that are “AI Native”—they’re growing up with this stuff baked into how they think about solving problems.

I thought I’d share the practical examples of what I built after listening, in case it sparks some ideas for you, dear reader, as well.

On personalized software

I’ve heard over and over that this is the time of highly personalized software. You can build software with a total user base of one: yourself.

When thinking up examples in the past I’ve always thought “well, in case I want a calculator but need it to be cat-themed, then that’s the perfect use case for personalized software!” Tada, there it is!

That’s not exactly the most practical example though. I really love cats, but no matter how much I love them, I’m still happy to settle with the pre-existing calculator app on my phone or computer instead of creating a custom one just for myself when I need to calculate something.

However, building my own productivity tool made this personalized software thing actually click for me. I’m building a tool that works exactly the way I want it to work. I can create “Focus” modes on tasks that take everything else off of the screen except for a pomodoro timer and the notes for the task, so that I can do a single thing at a time. I can build an app in the background that will automatically sync my GitHub notifications every 15 minutes and will only show the ones that I choose to show—say, a coworker on my list of coworkers marks a pull request in a specific codebase I care about as “ready for review.” THOSE are the notifications that I want, not every single pull request opened across all of GitHub.

Building personalized software is such a huge unlock. I think this is really the path towards those big productivity goals.

Command Center - my custom productivity tool

I’m constantly working on my own day-to-day productivity workflow. I’ve got an awful lot going on and the act of keeping all of those things organized feels like a whole job in and of itself.

In the past I’ve used the PARA method with Obsidian, but I found I wanted a change in the New Year. I wanted more customization options and an ability to add the notifications that I actually care about. I wanted to do more analytics on the amount of things that I get done each week, so that I can understand if I’m actually getting better and progressing towards my goals. I also wanted a tool that was fast, cheap, and pretty.

A lot of the things I wanted were highly custom to myself.

I am also very much a visual person. I’ve never been great at using the CLI for anything. I think a lot of folks have been able to build workflows for themselves using CLI-based tools, but I just can’t do it… I need a window!

For that reason building my own thing made a lot of sense.

Before having Command Center to help me organize my thinking and get a visual of what my ACTUAL priorities are—and to have a quick capture for me to jot down all those ideas that pop into my head on a whim and get them out of my brain—I was floating listlessly between tools, feeling like I could barely grasp the entirety of what I have going on.

Now I have a personalized tool that works exactly for me, that Copilot is already built into so that I can iterate on ideas in the moment as I’m working on something! I can connect all of the disparate apps where the state of things I have going on exist, like my calendar, or a google doc, or whatever, and be able to bring all of that context together into one place through MCP. It truly does feel like magic and I’m falling in love with AI all over again.

Oh, and the icon is cat themed. Because of course it is.

How I built Command Center (in a day)

I went from brain dump to a working prototype that replaced my Obsidian setup in a single day.

I’d been dreaming up this tool for about a week—sitting down and drawing out all of the things that I wanted this command center to do. Mostly to help me stay organized and keep track of my various scripts and agents that are in disparate locations, and to help me focus on what tasks I’m handling synchronously (like me being in an IDE doing the thing) vs. what I’m tracking and handling asynchronously (like assigning work to Copilot coding agent).

After that I opened up a speech-to-text tool and spent about 10 minutes describing all the things that I wanted, including things like the vibe I wanted to evoke—like staying in flow. I asked Claude to take all of that and devise a spec that I could use to build the tool. I then asked Claude to show me three different designs, and picked the one that I thought I liked the most. Then I asked Claude to write detailed issues that I could give to Copilot to complete the work, and slowly fed those issues to Copilot throughout a day.

There’s a recent post by Andrej Karpathy that really speaks to me:

“Given the latest lift in LLM coding capability, like many others I rapidly went from about 80% manual+autocomplete coding and 20% agents in November to 80% agent coding and 20% edits+touchups in December. i.e. I really am mostly programming in English now, a bit sheepishly telling the LLM what code to write… in words.” — Andrej Karpathy

I feel like I’ve made the same switch in the past several months and it’s weird. My brain doesn’t feel wired to rethink my entire view of technology quite this quickly.

On vibe coding

I think that vibe coding is useful in primarily one place: building a tool for you where you don’t care about the learning process of how to build that tool. As a software engineer who has been in the industry for a while, I think that’s perfectly fine to do! If you’re building a tool where the n is 1—you—then absolutely vibe code it. Make sure it works exactly the way you want it to, and who really cares what the code is doing under the hood. (After all, you’re using version control, right? Worst case if the agent goes crazy and burns it all down—something I’ve heard rumored but never actually seen myself—you can go back to the last working version!)

I’ve become more comfortable with vibe coding than I ever have in the past. The tools got better. Like, way better. I think as I’m slowly giving larger and larger tasks and seeing them be done correctly I’m getting more comfortable with it.

While I say “vibe coding,” anything actually in production for work still gets a lot of scrutiny. Working on a few of my own projects from scratch helped me get comfortable. When it’s a thing I’m building from scratch just for me, I have a lot more freedom to introduce bugs and have it be okay.

On MCP

I’ve known of the magic involved in MCP for a while but have never really gone through the act of setting it up. It’s like… really easy to set up.

I get the basics: a Model Context Protocol server is basically a way for an agent to talk to another application easily. It gives more tools to the agent, like access to your calendar.

This becomes very helpful for building your own productivity system. I can hook up the MCP for my own calendar provider and create a script that will sync my calendar events from my work day into my personal productivity system. It’s like my own system has a calendar integration without actually having to build one myself!

On agents

I’ve been experimenting with agents in some pretty interesting ways recently. The one I’m most excited about is a workflow that will automatically assign backlog issues to me and to Copilot, so that I’ll get pinged to take those issues through to completion. It’s been an incredibly cool tool for chipping away at tech debt without it feeling like a monumental effort.

I’ve also been playing around with other automations—like triggering research on podcast guests automatically when they sign up for the podcast, or taking care of some minor tasks when it comes to editing. None of these are huge, earth-shattering things on their own. But together they add up to a lot of time saved and are sparking even more ideas for me on where I could automate more things with agents.

On skills

Skills have been useful for things that I have to do occasionally—tasks that happen often enough to warrant automation, but not frequently enough that the process stays fresh in my mind.

For example, I recently built a “skill” that will help with automating a task we occasionally have to do to help folks integrate a new SKU in billing. I provided Copilot some past issues that detailed the process and just did a little transcript brain dump of everything I was thinking of. I gave all that to Copilot to build an issue template and a GitHub Action to power it, and within about an hour I had an entire automation to handle this use case. It happens semi-frequently—every few months—but not frequently enough that it’s “easy” or research-free. Now it kind of is.

In some cases it makes sense to have instructions that apply to everything, but skills shine for those occasional tasks where you’d otherwise spend 30 minutes remembering how you did it last time.

On being pragmatic

Here’s where I have to pump the brakes a little, because I think it’s important to be extremely pragmatic about all of this. While AI is incredibly useful in many contexts, it isn’t for everything. You have to be extremely careful, accept “good enough,” and be okay with saying “well that didn’t work, but that’s fine, I’ll try something else.”

I’ve been accepting “good enough” quite a bit, especially for fixing failing tests or fixing TypeScript errors. As long as things are compiling and make a reasonable amount of sense, I’m not looking at it too closely.

And things don’t always work! Sometimes I have to prompt the coding agent 3-4 times to really get a CI failure fixed, and sometimes it’s just easier for me to go into an IDE and fix it myself. That’s fine. It’s all part of the process.

On safety

I also want to be clear about boundaries. Don’t give an agent access to your entire life. The safety just isn’t there yet.

For example, I’m not about to give an agent access to any smart home things in my house. Or access to any significant amount of funds. I still want to make final purchasing decisions.

Although, let’s see what I say in a year given how profoundly my opinions have changed in the last year.

The takeaway

Times are changing, like really fast. ClawdBot went from ClawdBot to MoltBot to OpenClawd, all in just the last week!

If you thought that the tools just weren’t good enough in the past, well, they are now. If you’re in the software space it’s increasingly important to jump on the AI bandwagon if you haven’t yet. And also crucial to make sure you continue being pragmatic with it. Experiment and see what you can accomplish with these tools—but keep your eyes open and your boundaries clear.

I hate the term “AI-pilled,” but here we are. I guess I’m AI-pilled now.