Pre-grade student work against a rubric with Claude, then review it yourself
Idea + starter prompt, not run by us yet
For: Teachers and trainers who mark many written submissions
A lecturer used a script to pull submissions from Canvas, extract the text and send it with a YAML rubric to Claude. Claude returned a score per criterion, a short comment and a flag. He checked it all in a CSV before confirming the upload. On one lab, 13 of 22 submissions were flagged.
Steps
- Set up Canvas API access the way your school allows. Keep the token to yourself and never paste it into a chat.
- Write your rubric in a YAML file with each criterion. Include sample answers worth a 90, because the author's scores didn't match his own grading.
- Run a few anonymised test submissions first and compare Claude's scores with yours.
- Read the CSV and every flag. In the author's run, flags were mostly suspected AI writing or thin evidence. Treat a flag as a reason to look, not a verdict.
- Upload grades only after you confirm. Re-runs call the API again because responses aren't cached.
Starter prompt
Help me set up a rubric based grader for written assignments. My rubric is in [rubric.yaml]. Work only on [3] test submissions that I have anonymised, and remove names and student IDs before reading them. First show me your plan and the exact output format: a score per criterion, a short comment and a flag with the reason. Write results to a CSV called [results.csv]. Don't upload anything to Canvas and don't change any grade. Flag work that looks AI-written or has thin evidence, say how sure you are, and never treat a flag as a verdict. Student data is private, so don't copy it anywhere else. Never ask me for my Canvas token or any passwords.
Original post: vitraag.com · found via agentusecases.ai. Written up in our own words. We haven't run it, so test it on made-up data first.