🛠️ Doing It Right, Finally

There was something quietly satisfying about today’s progress, like a knot loosening. The work itself wasn’t glamorous. I was mostly laying down the bridge between frontend and backend for session management. But I'm doing it properly.

Part of the reason it felt like a grind before is that I’d taken shortcuts. The session stuff worked, sort of, but it wasn’t hardened. It definitely wasn’t something I’d want to repeat or trust long-term. So I threw it out, started from first principles, and am building it up again: hardened requirements, clean plan, clear prompts. I haven’t even run them yet, but the setup feels solid. And that alone let my brain unclench a bit.

Of course, I was also fighting model outages all day. OpenAI launched GPT 5.0, and naturally, that meant spikes in usage and flakiness — models disappearing mid-stream, requests failing, the usual chaos. But weirdly, even that didn’t throw me off. The new model is good, and maybe I’m just getting used to these waves.

There’s no code written yet. But the plans are real. And this time, I didn’t rush past the part that matters.

Also, there’s something satisfying about how much more rigor I’m using now. Tools, procedures, architectural hygiene… I couldn’t have done any of this before the new llm-driven workflow. I didn't know to do it, and if I got a inkling of a new topic (like linting or e2e testing suites), I just wouldn't have gone down those rabbit holes. It was simply too time-expensive to do all that stuff, when I had to manually do it myself.

I think when I was feeling frazzled, it was because I was staring at how far away the finished app still is — something I could demo, or even use. And yeah, in a linear sense, it is further away. But it’s further away because I’m identifying real requirements. Things that didn’t even exist before — like frontend-backend contracts. And these aren't just useless, academic, "shoulds" anymore.

With the llm-driven workflow I'm building into Vulcan Forge, identifying these requirements means I can get them handled and integrated within just a day or two. That’s not scope creep. It’s the fractal nature of real progress: zoom in and discover more of what’s true.

This is the big advantage of being able work at this level of abstraction. The clarity I'm gaining is now actionable. I can actually learn enough, create a plan, and have the LLM implement it in such a short time that finding new requirements is really preventing the build up of tech debt to be paid later (with interest).

Even redoing session management — which felt like a derailment yesterday — now feels like solidifying the foundation. Auth is core to the whole system. If I get it right now, it’s a permanent asset. It’ll compound. In the process, I also learn more about what Vulcan Forge needs to do, so it's a win on multiple levels.