Community-built Claude skill prunes useless UI copy, works surprisingly well with Opus

floguo · x · 2026-11-10

Developer chaotictransfem complained that Claude (Opus) and other models keep adding verbose, self-aggrandizing labels in web apps; the community responded with a clean-copy Claude Skill that automatically cuts this kind of copy, and it reportedly works surprisingly well with Opus. The skill is public.
Its core claim: bad copy is usually copy that shouldn't be on screen at all, not badly written copy. Before editing tone, ask whether the string should exist, then walk a deletion ladder:
Would the user ever have wondered about this unprompted? Answering unasked questions plants doubts (e.g. noting "preview rendered at exact profile size" under an avatar cropper implies it might not be).
Does the label need to exist? Keep only when the value is meaningless without it ("Retry limit" needs no "how many times we retry"; a bare "30" does need "seconds").
Is it pre-empting a failure the user would never have considered, like "never wipes the folder like the old cancel did" — scar tissue from a fixed bug rather than a description of behavior?
It's a reusable prompt-engineering pattern: encode editorial judgment about what to delete as explicit rules the model applies to any user-facing string.

More from coding & agent