Size is a prompt

Claim: past a length, a file is asked to state its role; size alone never decides a split.

• Grade: convention, untested. The belief that smaller files have fewer defects is not supported.
• Evidence: against (file length).
• Check: logical-lines: past 80 statements, a judgment question; past 500 lines, a review.
• Reader styles: novice lowers the trigger; expert raises it.

What it asks

• Count statements, not layout: a sentence, a list item, a table row or a code block each count one; headings, blank lines and wrapping count zero. Re-cutting a paragraph into bullets changes nothing.
• Past about 80 statements, state the file's role in one sentence (one role per file). If it holds, the file stays.
• Past about 500 lines, review the file: split it only into pieces that change separately and that each stay whole.

Why

The trigger costs one sentence and catches files that grew a second role. The thresholds are conventions, tagged PROVISIONAL in checks/config.toml.

Limits

No size threshold has held up for defects: defect-proneness grows more slowly than size, so smaller modules carry proportionally more defects (Koru 2008; Koru 2010a); the shape of the curve is contested (Syer 2015). Splitting keeps total size, and smaller modules are proportionally more coupled (Koru 2010b). Over-split code becomes entangled: one piece cannot be read without the other (Ousterhout 2025). All these studies measure code, not prose.