Eventually I realized why and explained, "you know, I'm really just going to do a web search for what you just asked me, and maybe a couple more until I have a decent answer and then give you that answer. Let me show you how I would go about that".

From then on, they started getting into the habit of doing that for themselves. I think now with LLMs, they've kept the habit, but the LLM gives a more complete answer with fewer steps so it becomes the default. I think the magic of AI is two-fold (well, more than two, but two bullets for this conversation).

1. You don't have to "query". You can just braindump, and it'll build a context and figure out what you're looking for

2. It's conversational, so instead of filtering and tweaking results from the first query, your second "query" builds on top of the context from the first question, and you get a stronger result as the conversation continues.

https://www.youtube.com/watch?v=CHYxjpYep_M

Sometimes I daydream about a "solo mode" where the timings on lights are tighter and my music can follow me around the house when I'm up at night and nobody else is. But most times I'm trying to find the get-out-of-way averages that keep everyone happy.

Some things work great: Automated lights everywhere. Automated dimming of lights at night or sunset or whatever. Notifications when the laundry is done, or the cat litter is ready to be changed, or someone is at the door, or the garage door has been left open - all great. What music to play in what room at any time? Always changes. When to "dim all the lights" because Plex started a movie? But my son is building Legos in the dining room, and my wife is knitting and needs the couch light on. Sometimes I want it, but not every time.

For those things having a single button press is still a huge win over opening multiple apps and getting the right things set the right way for each participant.

After wiring my whole house with Ethernet and ceiling speakers, and now dealing with a couple leaky pipes and several problems from previous owners, I'm considering ways to make these things easily accessible/replaceable while keeping an eye toward aesthetics.

I've patched quite a bit of drywall, and I'm about mediocre at it. But it seems so silly and unnecessary to me.

Everything else in this world that requires maintenance comes with access panels and other means of easy access. In our living spaces, some of which should ideally last tens of years (mine is from the 1890s), we seal it all away.

Drywall is manageable and cheap, I agree. But it's more painful than it should be for something that _will_ require maintenance.

Pipes will fail. Wires will fail. Ducts will fail. Maybe not in 5 years, but over the span of 20, they will. Why make them so frustratingly inaccessible?

One very important thing to keep in mind is context management. Every time your agent reads a file, searches documentation, answers a question, writes a file, or otherwise iterates on a problem, the context will grow. The larger the context gets, the dumber the responses. It will basically start forgetting earlier parts of the conversation. To be explicit about this, I've disabled "auto-compact" in claude code and when I see a warning that it's getting too big, I cut things off, maybe ask the agent to commit, or write a summary, and then /compact or /clear. It's important to figure out the context limits of the model you're using and stay comfortably within them.

Next, I generally treat the agent like a mid-level engineer who answers to me. That is to say, I do not try to convince it to code like I do, instead I treat it like a member on my team. When I'm on a team, we stick to standards and use tools like prettier, etc to keep the code in shape. My personal preferences go out the window, unless there's solid technical reason for others to follow them.

With that out of the way, the general loop is to plan with the agent, spec the work to be done, let the agent do the work, review, and repeat. To start, I converse with the agent directly. I'm not writing a spec, I'm discussing the problem with the agent and asking the agent to write the spec. We review, and discuss, and once our decisions are aligned and documented, I'll ask it to break down how it would implement the plan we've agreed upon.

From there I'll keep the context size in mind. If implementation is a multi-hour endeavor, I'll work with the agent to break down the problem into pieces that should ideally fit into the context window. Otherwise, by this point the agent will have asked me "would you like me to go ahead and get started?" and I'll let it get started

Once it's done, I'll ask it to run lint, typechecks, automated testing, do a code review of what's in the current git workspace, compare the changes to the spec, do my own code reviews, run it myself, whatever is needed to make sure what was written solves the problem.

In general, I'd say it's a bad idea to just let the agent go off on its own with a giant task. It should be iterative and communicative. If the task is too big, it WILL take shortcuts. You can probably get an agent to rewrite your whole codebase with a big fancy prompt and a few markdown files. But if you're not part of the process, there's a good chance it'll create a serious mess.

For what you're doing, I would likely like ask the agent to read the mega python file and explain it to me. Then I would discuss what it missed or got wrong and add additional context and explain what needs to be done. Then I would ask it if it has any suggestions for how we should break it into submodules. If the plan looks good, run with it. If not, explain what you're going for and then ask how it would go about extracting the first submodule. If the plan looks good, ask it to write tests, let it extract the submodule, let it run the tests, review the results, do your own code review, tweak the formatting, Goto 10.

I had guessed it was stopped because it came to an unshoveled portion of the sidewalk. If it can't traverse that, it's not made for this city

I'm not fundamentally mad as these bots. But if they don't figure out how to make them work with other pedestrians, then I'm going to start cheering on any vandalism delivered upon them.

The ACA allowed me to get insurance for the first time since I'd left home several years before. I knew lots of other freelancers at the time who were in the same boat.

Of course in the following years, insurers found plenty of loopholes to increase prices significantly year over year - and this is why leaving the middlemen in the middle was a TERRIBLE choice - but at the very least the quality of those plans still has a reasonable low bar.

I still find myself on the ACA from time to time. I can't afford it. But the plans are still significantly better and thus more affordable than what was available before.

* Checking it closely myself, which sometimes takes just as long as it would have taken me to implement it in the first-place, with just about as much cognitive load since I now have to understand something I didn't write

* OR automating the checking by pouring on more AI, and that takes just as long or longer than it would have taken me to check it closely myself. Especially in cases where suddenly 1/3 of automated tests are failing and it either needs to find the underlying system it broke or iterate through all the tests and fix them.

Doing this iteratively has made the overall process for an app I'm trying to implement 100% using LLMs to take at least 3x longer than I would have built it myself. That said, it's unclear I would have kept building this app without using these tools. The process has kept me in the game - so there's definitely some value there that offsets the longer implementation time.

We didn't have much of a relationship. He had friends, but never close ones. He was weirdly mean or weirdly seclusive or weirdly awkward at times - and also incredibly intelligent and occasionally gracious and hilarious.

After he passed, I wondered if he might have been somewhere on the spectrum - but his peculiarities were simply ignored. A poor boy, in a poor urban neighborhood, with a dead father, being raised by an immigrant mother and immigrant siblings doesn't get diagnosed with much of anything - if they see doctors at all. And hey, he had a near photographic memory, and did great in school, so what's there to worry about?

It's always been "how he was", and that's probably ok, but I do wonder if he would have had a better or somehow different life if he knew more about _why_ he was the way he was.