Re:Let's Talk about LLMs
by hombre_fatal on Tuesday May 05, 2026 @12:10PM
(#48024474)
Oh, I see. I'll grant whatever you take vibe-code to mean since that seems to be the hang-up -- vibe-code prob suggests there's no process at all.
My point is that the planning phase and implementing phase are basically unsupervised, and all the work goes into the planning phase.
Yet I've noticed that over time, I'm not even needed in the planning phase because a simple revision loop on a plan file produces a really good plan. My role is mostly to decide what the agents should do next and driving the revision loop by hand (mostly because it's the best place for me to follow what's happening).
Over the last few months, I've been working on a polished luks + btrfs tool for managing a raid1 disk pool. It's the most complex thing I've built with LLMs.
I've found Claude Code (Opus 4.6/4.7) and Codex (GPT-5.4/5.5) really good at coming up with high-level architecture and then implementation spec. For example, they came up with a systemd state machine which is above and beyond what I would have polished had I built this before LLMs, and harder than I thought to get right had I done it myself.
I have various tricks to get these good results, like providing the source code to the linux kernel, systemd, btrfs, cryptsetup, etc. so that the agents never have to rely on their training data. But this is why I'm pushing back on the claim that LLMs can't even spec out an auth system.
My point is that the planning phase and implementing phase are basically unsupervised, and all the work goes into the planning phase.
Yet I've noticed that over time, I'm not even needed in the planning phase because a simple revision loop on a plan file produces a really good plan. My role is mostly to decide what the agents should do next and driving the revision loop by hand (mostly because it's the best place for me to follow what's happening).
Over the last few months, I've been working on a polished luks + btrfs tool for managing a raid1 disk pool. It's the most complex thing I've built with LLMs.
I've found Claude Code (Opus 4.6/4.7) and Codex (GPT-5.4/5.5) really good at coming up with high-level architecture and then implementation spec. For example, they came up with a systemd state machine which is above and beyond what I would have polished had I built this before LLMs, and harder than I thought to get right had I done it myself.
I have various tricks to get these good results, like providing the source code to the linux kernel, systemd, btrfs, cryptsetup, etc. so that the agents never have to rely on their training data. But this is why I'm pushing back on the claim that LLMs can't even spec out an auth system.