How I use AI

I wrote this in response to a Hacker News question.

Harness

For the past year or so I've been a Claude Stan. However, their recent availability issues combined with the Fable clown show, has caused me to consider other providers. My current daily driver is Grok 4.5. It's easily 2x (or more) faster than Opus 4.8, and feels just as intelligent (based on vibes: I find myself swearing less at Grok than I do at Opus).

Daily Work

While the question asked for "normal SWE work" I'm probably not exactly normal here. I own my own software agency, and so I do a mix of normal SWE stuff, exploratory fractional CTO stuff, and a bunch of scattered client management stuff. Thankfully, most of my day still involves shipping code.

I'm a terminal nerd, so everything I do is from a CLI. I don't use cursor. When I need to hand edit files, I use NeoVim or RubyMine.

Meeting Notes

I love transcription. Most meetings are not valuable beyond "What did we say we would do?" LLMs are incredibly good at extracting that information. Bonus points if your linear MCP lets you throw that directly into a triage queue. Extra credit if you say, "Ask me to clarify anything that's unclear"—you'll normally get better tickets this way.

Exploring visual designs

Something I've been doing a lot more recently when working on UI overhauls or new features: hammer out the entire design in pure HTML/CSS to get a feel for the proper visual constraints. Last week I had Grok create 4 different designs for a screen one of my clients wanted to improve. By listing a few constraints we wanted the design to satisfy we ended up with 4 designs, none of which were exactly right, but allowed us to identify the elements we did and did not like. This is way faster than having a designer step in and allows the stakeholders to visually hold the solution in their head before we commit to anything. It takes about 15-60 minutes depending on how much you nitpick/refine the designs the agent gives you.

Coding

Perhaps obvious, but I use agents to code. I'd wager 98% of my code is written with AI. Rarely do I find it more useful to go tweak code myself, however, I do read almost everything that's written. This has forced me to develop some interesting heuristics:

1. If I can't scan the code and understand what's going on—it's too complex. Send back to the agent for simplification. (We write a lot of Rails, so I have a /37signals-review command that has the agent review the code and ensure 37signals would be proud of it. If not, change it.)
2. If the code arrives at a solution I didn't anticipate, I slow down and scrutinize. Many times the AI will design a better solution than I would've, but I'm always skeptical.

Planning

AI is just as helpful (if not moreso) as talking to another human if your intent is to ask questions to push your thinking. Many times I'll start a conversation like:

I have XYZ idea, poke holes in it to see if it's worth doing.

Then, I don't have to think critically about generating questions I only have to think critically about the answers. With my limited brain, that feels useful.

What I don't use

Here's a list of features that I haven't yet found a use for, in no particular order:

• Loops (of the Ralph Wiggum nature, or otherwise)
• Goals
• Subagents (with one exception—I did have a feature once that neatly subdivided into 3-4 phases and I had an agent work on each phase in-tandem, but that's suuuuper rare, and it's usually obvious when that would be helpful.)
• Advisor/Coordinator modes

First Principles

The question specifically noted wanting to know about "first principles", which I take to mean: How do you think about this tool and it's role in your workflows?

First of all, AI is an exoskeleton, not an autonomous robot. Despite the fact that I unleash AI in "auto-accept" mode most of the time, I'm only delegating the last leg of the race to the robot. In fact, the generation of code isn't particularly valuable. AI exists to make me better, faster, stronger, etc—not to autonomously solve my problems. I frequently need to remind myself that less is more, just because we can fix this problem in thirty minutes doesn't mean I should fix that problem (now or ever!) It's required more discipline of me, not less, and I think that's a good thing.

Next, I believe that writing worth having another human read is writing that another human should produce.¹ When I produce client updates, it's all written by hand. I may have AI produce a first draft so that I don't have to dig through 30 tickets to gather context, but I always make sure what they read is something I would've reasonably written (including editing for voice, tone, etc.) This is related to my last post about client experience—it's more powerful than ever to engage in unreasonable hospitality.

There's definitely more, but that's all I can think of for now.

Last Updated: 8:20am Sunday July 19th, 2026