viktor's chronicles, notes and rumbles
A friend wrote a piece about starcraft and AI, and it really resonated with a lot of people. I was thinking about search in claude today and asked it to build a skill to use mdsearch and you can ask it how and when it is going to use it and claude just told me, and that feels so magical, in a sense this is main source of pleasure working with AI agents for me, ability to ask why they did something and get, not always real, but answer. And now you engineer turned tech lead, EM, PM and starcraft player trying to burn your company sponsored tokens as fast as you can.
And I do not think there is really a way back. From moving physical wires to punchcards to byte codes, and then from assembly to C and all the STL baggage and ultimately SQL there was no way back.
For years SQL was the highest level language we used routinely, because you started expressing what you want the program to do, not how to do it. Of course you lose control over execution — and that is why it is so awkward to use or not use an index or define the order of joins — all of this by design. And going from AI autocomplete to console Claude Code feels like a natural next step for me. Reason with Claude about what to do and watch it figure out how to do it, often poorly but passable, look up libraries and read logs without switching tool, all in the same context.
And now with accidental complexity and mental burden of installing dependencies or fixing formatting to make your linter happy, you have more space to work on that flight equation, or make entire screen render in 8 microseconds in terminal.
Taking CS classes I was always confused why ergonomics was not a value. I think the first time I heard anyone talk about this was Lou Grinzo in the Zen of Windows 95 Programming, and then the Ruby and Rails community. I always felt like most of the problems, after we got hundred megabyte hard drives were accidental, and a lack of focus in that area — though I admit this might be survivor bias. Mitchell Hashimoto is a great example of working on a problem that at once was what he enjoyed solving, was a huge pain and was not a necessity.
And yet the whole programming community kept going in circles with the same ideas — CORBA, SOA, REST APIs, microservices, modular monoliths. Debating tests or no tests. Spaces and tabs.
Although there is no surprise – number of engineers doubles and doubles, we had half of the people in the industry with a year or less of work experience, knowledge do not accumulate, we only started doing pair programming as discipline recently.
The first answer that I know of at breaking this was Emacs — you write your code in lisp and you program your editor in lisp, magical , but not everyone read and loved SICP, and it stayed niche, and also it could never tell you why your macro did not work.
But Claude Code is the first tool that is both the next step on the abstraction ladder, like you have better SQL, natural language and for all human knowledge. And it is genuinely ergonomic, as you already spend all your time in terminal and irc. You are now in a workshop where making new tools for your specific problems has the same texture as doing the work itself. And for the first time, an inquiring mind can ask a question to do deep research on how people solved a problem before, getting the accumulated institutional memory of what people actually tried and what broke, queryable when you need it.
I am not even sure it will exist in its current form two months from now with the speed things are moving. But even if it is not the claude, pandora box is open and we know what is possible. This is almost replicator already.