Open navigation
Welcome
Your task
Task workflow
The task form
Criterion guide
What good looks like
Rules of thumb
Common errors
Rubric rules
Atomic, not bundled, not stacked
No overlap, always consistent
Deterministic criteria
Cover every requirement
Writing new criteria
Editing the draft safely
Good and bad rubric, side by side
Pre-submit checklist
Finish the review
Edit rationale
Quality expectations
Check your understanding
For reviewers
How to review a task
Keep an eye on the details
Fix and approve, or send back
Writing feedback
Reviewer checklist
Pay and logistics
Get help
FAQ
Changelog
Reviewer checklist
Work through each item.
Checklist
0 of 20 complete
I read the final request, the whole conversation and every image myself.
I wrote my own picture of what the user wants before reading the fellow's rubric.
Every claim about an image matches the image.
Every number, version, name and path matches the user's messages exactly.
No requirement comes only from an earlier AI reply.
Every goal, limit and requested step from the user has its own row.
The most tempting wrong answers have penalties, written as the mistake.
Every row is atomic: nothing bundled, nothing stacked.
No single mistake fails or triggers two rows; no penalty mirrors a critical; no bonus mirrors a penalty.
No two rows contradict each other, and the summary matches the critical rows.
No judgment words; numbers have a tolerance; conditional rows say what happens otherwise.
Nothing is required, and no bonus offered, for something the user did not ask for.
Every critical passes the leave-it-out test; every bonus is optional and in scope.
Image numbers list only the images a grader needs.
A different valid approach and a short complete answer both pass; a plausible wrong answer fails.
Technical claims are checked against official documentation.
Draft IDs are kept; new rows are new_N; deleted rows are explained in the rationale.
Every change log matches the real edit; the outcome matches what was done.
If I fixed things myself, I updated the change logs and added a review line to the rationale.
I scored the fellow's submission before my edits, and I approved only after re-checking the whole rubric.
Writing feedback
Pay and logistics