squint: the skill told the reviewer to read thirty-four screenshots

Testing it end to end found three things the text got wrong. Reading every
shot in review mode is not possible -- four full-page screenshots is
already a lot of context -- so it now says to pick a handful and name which,
and to offer to go deeper rather than silently reading twenty more. A
design guide is mostly about how to build the thing, so it points at the
sections describing what a rendered page should look like. And it says to
report finding nothing, because a review that invents a finding to justify
itself is worse than one that says the screens look right.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X1aAjLD4wvan3wsmkHPPh3

James Millscommitted 1 month agoparent 0127645c5251422a302a16ef5c7ee652fe0251597a32c60Browse files