Most rewrites serve the engineer, not the business

There's a confession buried in this essay that a lot of engineers will recognize with a wince. The author spent weeks waking up at 4am to rewrite a CakePHP app into Laravel, on his own time, unasked. Nobody requested it. The app worked fine. The rewrite changed nothing for users or the business, and improved exactly one thing: how the code felt to him.

That's the pattern he's naming, and it's worth sitting with. Production code that's survived years of real traffic is a ledger of every bug someone already found and fixed. Joel Spolsky said as much back in 2000, calling the full rewrite the worst strategic mistake a software company can make, because you don't just throw away ugly code, you throw away the fixes hiding inside it. The author's own examples back this up: a Perl system over a decade old, a homegrown PHP CMS about the same age, both untouched, both still earning their keep quietly every day nobody rewrote them.

The deeper trap is mistaking unfamiliarity for brokenness. He admits he'd never actually learned CakePHP before deciding it was garbage. That's the tell: when you don't understand a system, every choice in it looks like a mistake, because you can't see the constraints that shaped it. The fix he proposes isn't a freeze on all rewrites, though. Some debt is real. End-of-life runtimes with open CVEs, a bus-factor of one walking out the door, feature costs tripling because of bad architecture, a business need the code was never built to support, these are forcing functions with numbers attached. If you can point to a CVE or a velocity tax or a missed deal, you have a case. If all you have is a feeling that the code offends you, you have a preference.

Then there's the AI wrinkle, and this is where the piece gets sharper. The new argument going around is that AI makes rewrites cheap, since an agent can port a module in an afternoon. But the author points out that typing was never the expensive part. Rediscovery was. The reason a weird timeout exists lives in a 2021 Slack thread nobody linked to a postmortem, and the agent has no access to that. It writes something clean and plausible with the scar tissue quietly removed, and then the team meets the old bugs again, minus the context the original fix carried. Cheap generation doesn't shrink the rewrite problem, he argues. It just makes it easier to start and harder to stop, because in two years someone opens the AI-written version, doesn't understand why it's shaped that way, and wants to rewrite it too.