Software Engineering

I’ve read a lot of observations from people signaling a similar epiphany I’ve been having for a while - that well-used AI allows you to profit from useful bits of atypical brain wiring while lowering friction of problematic ones. Yet, the common approach focusing mostly on “what” at the cost of “why” and “how” did not sit well with me. Founder-mindset and clockmaker-mindset are different. Now I feel that I’ve built a flow giving me similar benefits, while keeping me honest to what I value. ...

Waterfall-in-disguise’s return is just around the corner. If we allow it, we will be even more miserable than we already are with “scrum” cargo-cult. For the second time in the last year I will be writing minimal OAuth2 auth server (why that is material for another note). Creating & refining initial spec took a full day of work so far - before even one LOC was created. The model’s ability to spot underspecified requirements or contradictions with chosen RFC baseline is a godsend when working with a sensitive domain like security. Makes good spec cheap, not resurrecting waterfall by itself. Still, three serious problems follow: ...

I’m prone to being overly enthusiastic when something seems to fit perfectly puzzles I try to solve. Right now I’m in that state exactly. Ubiquitous language is the answer to so many pains of agentic coding. Part of my current experiment vocabulary started as “Workflow”, then it gave birth to sibling “Component”, then “Fabric” was extracted from “Component” when its meaning got refined, and then “Fabric” revealed “Substrate” underneath. Each of the evolving terms guiding both me and the agents in the correct direction, constraining what conceptually makes sense within the bigger picture defined by vocabulary. ...

I’m starting to think about “grilling” (in the context of Matt Pocock’s skill framework) not only as design/decision-recording sessions, because the core benefit is how they expose how domain language clashes with your intent—requiring adaptation of one or the other. The definition of terms is your core pre-code-generation guardrail. The “world model” of LLMs is language itself, so it makes them pretty good at spotting language-, vocabulary-, and definition-level contradictions. Often better than we do, if you are not blessed with a tiny bit of autistic traits. ...

Adopting ubiquitous language (where agents are heavily responsible for keeping you, the codebase, and the vocabulary in alignment—otherwise, drift will kill you) greatly improves alignment between what you want and what agents deliver. There is still tactical cognitive debt. You have no idea about the details of delegated implementation, but that’s OK. You decided to be a Lieutenant General, not a Lieutenant, so you move divisions, not platoons. You note that a particular operation introduces side effects that break everything that follows. You can change the implementation to mitigate them; you can ask an agent to do it—it might be enough to move forward. Until a minute, hour, day, or month later, when you or an agent breaks the silent, tribal contract you made by fixing the implementation to be pure and idempotent. It won’t fly at scale. Shit, it won’t fly even in domain discovery mode for a non-trivial personal project. ...