Meh. Not everybody gets clicks on their blog.

this in fact is excellent articulated what I also strongly feel.

went from decades of oldschool crafting by hand and got good at it, then went into vibecoding and the experience of decades made me leverage agents significantly better than other peers, but it is also addicting and getting lazy, fast. to the point I type "git push" into claude regularly, not only because I am too lazy to switch my UI tab, but also because occasionally there is this bit of git friction with some local/remote states and claude/codex just resolve it instead of me getting distracted with it.

and now I essentially force my squirrel brain the opposite way again. everything new and exciting and MVP grade ideation happens with vibecoding in hours/days, shippable in isolation somehow so strictly limited blast radius, and then only the <10% of stuff that turn out valuable enough I go back to and rewrite/extract by hand, literally typing it out manually, _without_ autocomplete beyond what a lsp in zed gives me (no cursor magic, ...). AI is only used as a assistant to discuss solution options quickly like a coworker/mentor on the side, but even then I type the result manually.

And it feels like an alien now to do this but its getting better slowly and boy, even my handcrafted code gets better than ever, slow and deliberate, and also I switch to OCaml now for everything that matters, having learned it through LLMs for weeks. No more Node/Go/Rust for me, finally settled after like 20 years. So a quick slop of vibed experiments (react, node, go services, whatever fits best quickly), and the important stuff then slow careful crafted rock solid and lightning fast ocaml.

feels like this is the way. even if it hurts still, kinda like DOMS after going back to gym after long breaks.

i run a small insurance firm, 20 years in that domain, zero in software, and i recently vibe-coded my way into running agent infrastructure for the business, so i should stall at 80%. but I think that the judgment my project actually needed was mostly regarding insurance itself, and making sure the knowledge base it pulls from was sound, so I was able to use my experience on that side to produce what I hope is a useful product

the technical side is of course different and i don't have that judgment, so i try to get by with fresh-session reviews, testing, and keeping it on the simpler side because i know i can't audit the code myself. whether or not that proves to be a winning formula remains to be seen, but either way i think it's reasonable to say that some regular folk will stall at the 80% but others may figure out a way to power through it using caution and reason