1890
Up to and including 1880 the country had a frontier of settlement, but at present the unsettled area has been so broken into by isolated bodies of settlement that there can hardly be said to be a frontier line. In the discussion of its extent, its westward movement, etc., it can not, therefore, any longer have a place in the census reports.
Disclaimer: I am a shareholder in Anthropic. That position has absolutely no bearing on what follows.
For the past year, I've tried to timestamp my thoughts on using LLMs for software engineering every month or two — LLM as advance team, Scattered thoughts on LLM tools, Fable 5.1, and It's just a model, to name a few. I realized it was time to write another one, and I found myself with very little to say.
I have little to say because, since Opus 5.5 was released and calcified, I have done essentially no tinkering with my day-to-day work. I still use Conductor. I still, for the most part, use local environments — unless I'm kicking off a long-running task right as I'm about to head out, in which case I'll use Claude Code and its cloud sessions.I don't use Conductor's cloud offering because it still feels a little finicky. Every time I try to kick the tires on it again, I find myself having to restart the cloud worker or futz with something. I use Opus 5.5, exclusively.
And that's sort of it. I don't really manage my context; I don't write agent-specific files; I don't worry in the slightest about running into token limits. Opus 5.5 is the apotheosis of LLM engineering for me. It's smart, but not too smart. Fast, but not too fast. And I've spammed "use simplified technical English" enough times that the prose it spits back out at me is acceptably legible.
As far as I can tell, I'm not alone in this. Most of the folks on the team who have a wide berth to use whatever tool they want have landed in the same place.
You might wonder what this has meant downstream for our engineering culture — how my thinking about code review, technical debt, leverage, and so on has evolved. The answer, quite bluntly, is that it really hasn't.
Opus 5.5 is as close to a perfect model as I've ever used. And despite that — just like any good engineer — it gets some stuff wrong. What it doesn't get wrong, it might get correct, but only technically. And even when well-intentioned and explicitly prompted, it biases (again, just like any good engineer) towards point fixes over systemic ones, and towards addition over subtraction.
What's more, invoking Opus 5.5 does not improve my ability to invoke Opus 5.5, let alone do anything else. That improvement only comes from thought and review and consideration: from investing in internal tools (a task now made much easier), and from thinking hard about technical and product challenges.
This might not be the most galaxy-brained way to do my job right now, but I find myself the happiest I've ever been as an engineer. I am spending more time than ever before thinking hard about difficult problems — and at the same time, I still get the joy and sense of compounding aptitude that once came from spending hours writing code by hand.