How to review a task

Before reading the fellow's rows, write down in a few bullets: what the user wants from the final message, every limit the user stated in any turn, every step the user asked for, what each image actually shows, and the one or two most tempting wrong answers. This is your answer key. The fellow's rubric is judged against it, not against the draft.

Compare the conversation, the draft and the fellow's version. Sort each row: kept right (the draft was right and the fellow kept it), fixed (the fellow improved a weak draft row), broke (the fellow's change made a good row wrong, too narrow, or too loose), missed (a weak draft row the fellow kept), added (a new row; check it like any other). Give credit for fixes.

Atomic, not bundled, not stacked, no overlap with any other row, consistent with every other row, deterministic wording, in scope, grounded in the user. Then check the weight (leave-it-out test for every critical), the penalty form (written as the mistake), the dimension, the image numbers (only the images a grader needs) and that the change log matches the actual edit.

Every goal, limit and requested step from the user has its own row. The most tempting wrong answers have penalties. Deleted draft rows are explained in the rationale. The summary matches the critical rows. The outcome matches what was done.

A strong answer, a different valid approach, a short complete answer and a plausible wrong answer. The first three must pass; the last must fail a critical or trigger a penalty.

If a row depends on how a language, library or tool behaves, check the official documentation. Never use an earlier AI reply in the conversation as proof; those replies are often where the bug came from.

Fix and approve, or send back. See Fix and approve, or send back.