JetBrains ships a version-gated Go rule set for agents that write decade-old idioms

The hand-rolled loop is the more interesting of the two failure modes, because it is not ignorance. The magic number 13 in that scan is exactly the character count of "invalid_grant" ^[1], so whatever wrote it measured the needle and then built a bounds-checked walk over the string rather than call a function that has shipped in the standard library for the language's entire public life. It also left a comment justifying the choice ^[4]. That is frequency bias at its most expensive: not an absent feature, but a plausible rationale for avoiding one, which survives review because reviewers read comments as decisions rather than as output. The replacement was a single call to strings.Contains ^[5].

The design decision here worth copying is the version gate. `list` resolves the applicable rule set from go.mod, go.work, or the local toolchain ^[7], and the write-up's stated reason is blunt: without it an agent will cheerfully propose errors.AsType[T] and break a build running Go 1.24 ^[9]. Most house style guides fail in the other direction, describing a language version nobody's CI is on. This one refuses to describe features your own module cannot compile.

The content is thinner than the framing suggests, and the diff between two adjacent versions shows why. Moving from 1.21 to 1.22 adds five entries ^[10], roughly an eighth of what a 1.24 project sees ^[2], and most of them are one-for-one substitutions: a new stdlib call in place of an old one, a loop form in place of a copy. That is a changelog with opinions attached, which is a reasonable thing to hand a model, but it is not architecture advice.

Where `explain` earns its place is the caveat text. For cmp_or, the details note that all arguments are evaluated before the call ^[11], which is precisely the regression a mechanical rewrite introduces when the fallback branch is an expensive lookup. An agent given only the one-line summary makes that substitution everywhere. An agent given the explanation has a reason to decline it in one place out of ten, and knowing when a rule does not apply is the part of a style guide that usually goes unwritten.

Two limits are worth stating plainly. This is one developer's test of one file in their own project, a 1,039-line main.go ^[12], not a measured before-and-after across a codebase. And the tool is a CLI with two subcommands that print text when asked ^[6]; nothing described makes it fire during generation. The knowledge gap is now addressable. The habit of consulting it is still yours to enforce.