Built fast in Cursor. Now nobody can safely change it.

"No consistent pattern" is rarely one big mistake. It's several ways of fetching data in the same app: a raw request written by hand here, a thin custom wrapper around the same call there, a data-fetching library adopted for one screen and nowhere else. It's near-duplicate helper functions doing almost the same conversion with slightly different rounding rules. It's one route that throws on a bad request, another that returns null, and a third that logs the error and moves on.

This isn't sloppiness. Each suggestion Cursor offers is generated from the immediate context in front of it — the open file and whatever else gets pulled in — without a persistent memory of conventions the team settled on elsewhere in the codebase. It has no way to know a convention exists unless that convention happens to be visible in its current context.

On one file this barely matters. It compounds because accepting a suggestion is fast and each one is judged only against its own diff — does this function work, does this component render — not whether it matches what's already there. Nobody is asking the question a senior reviewer normally would: have we built this already. Skip that question often enough and you get architectural erosion: many individually defensible decisions, with nobody responsible for the whole, drifting the codebase away from any coherent design.

That's what makes it hard to catch in the moment. Each suggestion, on its own, looks fine: it compiles, it solves the immediate problem, it doesn't crash in the demo. The inconsistency only shows up in aggregate, once enough files do the same thing several different ways and nobody remembers which version is the real one.