earn-every-line

Earn every line

The model's default is to add: the helper the function did not need, the option nobody passes, the try/except around code that cannot fail, a comment per line saying what the line says. Anthropic's own docs admit the tendency of its models (the "Overeagerness" section, written for Opus 4.5/4.6: they "overengineer by creating extra files, adding unnecessary abstractions, or building in flexibility that wasn't requested"), and this library's unslop-code catalog corroborates it from complaint data (over-engineering verified at 7.8% of tell-naming complaints, narrating comments 8.5%). This skill is the write-time gate those numbers point at. Layer: leaf. Admitted by commission plus external evidence rather than the misses log — see Provenance — so candidate-tier until fires accumulate.

The rule

Every line serves someone who exists. Code serves a caller that exists; a comment serves a reader who learns something the code cannot tell them; flexibility serves a requirement someone actually stated. Name the caller, the reader, or the requirement — by name, not by category ("a future maintainer" is a category; "the retry loop in sync.ts" is a name). Cannot name one → it does not go in. The test is positional, so it travels: the check that is rigour at a system edge is noise three calls deep, a public API owes the docstring a module-private helper does not, and a report section or plan phase answers the same question — who needs this now?

Deletion gets the mirror test, and it outranks every licence below: name what is lost if this line goes — the caller, the reader, or the requirement it serves. Can name one → it stays. Concretely banned as "fixes": stripping validation at a trust boundary, deleting a comment that carries a why, golfing for line count, and citing YAGNI to skip tests or refactoring — Fowler's own exclusion: yagni "does not apply to effort to make the software easier to modify" and "is not a justification for neglecting the health of your code base."