Feedback on Student Writing¶
The Prompt This Guide Uses
Rubric Feedback Drafter
Drafts quote-anchored observations on de-identified student writing against your frozen rubric.
Have ready: your rubric or feedback criteria, written out, the assignment brief the students actually received, and the submissions, de-identified as step 1 shows.
Works with any capable assistant in the Tool Directory.
The Workflow¶
- De-identify first. Remove names, identification numbers, and identifying details (a clinical encounter described with its site, date, team, or an unusual detail can identify a student as surely as a name). The feedback is about the writing, not the writer, and leaving names out can help you read the work rather than the student, which is why this step comes first. Labels such as "Student A, Student B" keep your ability to map feedback back; then read each essay for the indirect details a find-and-replace misses.
- Freeze the criteria. Give the model your rubric (if it lives in your head, this task is the reason to write it down) and the assignment brief, have it restate the criteria as a numbered list, and correct it before any essay is read. The Rubric Feedback Drafter builds in this step and the next. Consistency across the stack is the whole point; a rubric that drifts mid-stack is worse than none.
- Run per essay, observations only. For each submission ask for: rubric-grounded observations, each anchored to a verbatim quote from the essay; the single highest-leverage improvement; and one thing done genuinely well, also quote-anchored. Ask it to leave out grades, scores, rankings, and comparisons between students (the Rubric Feedback Drafter already does). Measure each essay against the rubric rather than a classmate's essay: a peer exemplar pulls the feedback toward comparison, and the rubric is the standard the students were given. For a long stack, start a fresh chat every ten or so submissions, with the prompt and the confirmed criteria pasted again, so earlier essays stop shaping the feedback on later ones.
- Judge and rewrite. Read each observation set against the essay. Discard what is wrong or tone-deaf (some will be), keep what you would have found yourself on a good day, add what only you can know, and rewrite the keepers in your own voice.
- Return it as yours, disclosed as your syllabus says. How you disclose artificial intelligence (AI) assistance in feedback follows academic standards and the course and departmental guidelines the AI Responsible Use Policy points to; your syllabus AI statement is the place that decision lives, and students reasonably expect the same transparency about your AI use that you expect about theirs.
Good Practice for This Task¶
- This playbook is for formative feedback, and it leaves grades out on purpose: models drift across a stack and can praise or fault things the essay does not contain, and a grade is a judgment you need to be able to explain to the student. If you want to explore AI support for scoring, treat it as a pilot: compare its marks with your own across the stack, look for patterns that disadvantage particular groups of students, and share what you learn with the AI Committee.
- Read the observations as notes on this text, not on the student: the model saw one essay with no history, so its reading says little about the writer.
- If your course limits students' AI use, they will notice whether yours meets the same standard. Disclosing your own use the way your syllabus describes keeps that conversation honest in both directions.
About This Task¶
Give substantive, rubric-grounded formative feedback on a stack of student writing (reflections, essays, reports, case write-ups) in a fraction of the usual time, while the grade, and any judgment of the student, stay with you.
Where AI Helps, and Where It Hurts¶
AI is strong at the mechanical layer of feedback: applying your rubric consistently to the twentieth essay as to the first, spotting structural patterns (a missing counterargument, an unsupported claim, a conclusion that answers a different question), and offering alternative phrasings for points you want to make more kindly or more clearly. It is weak at everything that requires knowing the student (judging growth against their previous work, sensing what this particular writer can hear right now, and any decision that touches a grade), because it has none of that context. It also fails in a way specific to this task: models generate plausible-sounding praise and criticism that does not match the text in front of them, so every observation must carry a quote you can check. For a report or case write-up with references, the model can judge how a source is used in the argument, but not whether the source exists or says what the student claims, unless you give it the source; check the citations yourself.
One more failure mode is yours rather than the model's: feedback that arrives in a voice that is not yours teaches students that your feedback is not worth reading. The model drafts observations; the words that reach the student are yours.
From submission to feedback
De-identify first
names, IDs, traceable details out
Frozen rubric, observations only
each anchored to a quote
You judge, cut, and rewrite
in your voice
Feedback, disclosed
Grades stay with you; this pipeline drafts formative feedback.
De-identifying protects the student's record; keeping grades out of the pipeline keeps the judgment yours.
Related: Your Syllabus AI Statement, where your disclosure practice is written down, and the advisor note in Shelf and NBME Score Reports.
Content last reviewed September 1, 2026. Review dates are maintained as described on the About page.