Using Generative AI to Draft a Rubric

Most TAs don't write the assignment. They are handed one and asked to "make the rubric," or to fix a rubric the graders kept disagreeing on. A generative AI tool can help with both: drafting criteria and levels, listing the mistakes students are likely to make, sharpening descriptors, and writing practice answers for calibration. It cannot know what your instructor counts as correct, and its descriptors often sound precise without being so. This page gives you six prompts, a checklist for judging what comes back, and the rules for using AI here. If you are writing a new assignment and its rubric together, use Using Generative AI to Draft an Assignment instead; its prompt produces both.

Four rules.

Ask your instructor first, and use a tool the course or the university allows.
Never paste real student work, names or grades into an AI tool. That is student data, covered by FERPA. Use made-up answers instead (prompt 5 writes them).
Check every criterion and descriptor yourself, and try the rubric on real submissions before grading, as you would any rubric.
Say that AI helped, on the assignment or in the syllabus, as you would for an assignment.

Which comes first, the assignment or the rubric?

The learning outcomes, then both together. Decide what students should be able to do, then what evidence would show it (the rubric's criteria), then the task that produces that evidence. In practice the two are drafted side by side: every part of the assignment needs a criterion, and every criterion needs a part that produces it. The rubric is finished last, after you try it on real answers. When the assignment already exists, as it usually does for a TA, start from its outcomes and solution and build the rubric to match.

Prompt 1: a rubric for an existing assignment

The one you will use most. Paste the assignment and the solution where marked, and replace every [bracketed] part.

You are helping a teaching assistant write a grading rubric for an
assignment in a university computing course. Work as an experienced
instructor who writes rubrics that different graders apply the same way.

COURSE AND ASSIGNMENT
- Course and level: [e.g. COMPSCI 250, second-year discrete math]
- What students should be able to do after this assignment:
  1. [outcome, with a verb you could observe]
  2. [outcome]
- The assignment (paste it below the line):
---
[assignment text]
---
- The solution or answer key (paste it below the line):
---
[solution]
---

RUBRIC SETTINGS
- Rubric type: [analytic / holistic / checklist / single-point; or "recommend one"]
- Total points: [N]. Split across parts: [if known]
- Scoring: [add points for what is there / deduct from full marks]
- Number of levels per criterion: [e.g. 3 or 4]
- Graded by: [autograder / people / both]. If both, say what the
  autograder checks: [e.g. correctness via hidden tests]
- Results students may use without proof, or code they may reuse: [list]
- Course conventions to follow: [notation, definitions, style rules]

RULES FOR THE RUBRIC
- Every criterion must be something a grader can see in the work. No
  value words such as "good", "clear", "insightful" without saying what
  they look like.
- One idea per criterion.
- Levels must be clearly different from each other.
- Correct alternative solutions must be able to earn full credit.
- Name each expected mistake and its exact cost, and say whether it is
  charged once per problem or every time.
- Keep "attempted but incorrect" separate from "no attempt".

WHAT TO GIVE ME
Before writing, list your assumptions and any questions. Wait for my
answers. Then produce:
1. The rubric, as a table: criterion, points, level descriptors.
2. The 5 to 8 mistakes you expect most often, each mapped to a rubric
   item and its cost.
3. Two correct answers that differ from the solution, and how the rubric
   scores them.
4. Any criterion where two graders might still disagree, and how to
   make it sharper.
5. A short list of what I should check by hand.

Prompt 2: which kind of rubric?

I need to grade [describe the task in two sentences] for [N] students,
with [N] graders, in about [N] minutes per submission. The main purpose
is [a grade / feedback for revision / both]. Recommend a rubric type
(analytic, holistic, checklist or single-point), the number of levels,
and whether to add or deduct points. Give your reasons in a few
sentences, and say what I would lose with the next-best choice.

Prompt 3: tighten a rubric graders disagreed on

Describe the disagreements in your own words. Do not paste the student answers.

Here is a grading rubric (below the line). During calibration, graders
disagreed on these points:
- [criterion]: [what the disagreement was, e.g. "some gave 3, some 1,
  for a proof that skipped the base case but was otherwise right"]
- [criterion]: [disagreement]
---
[rubric]
---
For each disagreement, rewrite the descriptors so that two graders would
score the same work the same way. Add one short made-up anchor example
for each level of those criteria. Do not change criteria that caused no
disagreement. Show only what you changed.

Prompt 4: a version for students

Here is the rubric graders will use (below the line). Write a version to
share with students when the assignment is released. It should tell them
what each criterion looks for and how points are split, in plain words,
so they can check their own work before submitting. Do not reveal the
solution, the hidden test cases or the exact deductions for specific
mistakes. Where hidden tests exist, say what kinds of cases they cover.
---
[rubric]
---

Prompt 5: practice answers for calibration

Graders calibrate on shared answers. Real student work is best when your course allows it inside the course's own tools; it never goes into an AI tool. Made-up answers are a good way to practice before real work arrives.

Here is an assignment problem and its rubric (below the lines). Write
[6] made-up student answers to this problem for grader training: one
clearly at each level of the rubric, plus [2] borderline answers that
sit between two levels. Make them realistic, including the mistakes
students in [course and level] typically make. Do not label them. Then,
separately, give the score you would assign each one under the rubric,
with one sentence of reasoning per criterion.
---
[problem]
---
[rubric]
---

Each grader scores the answers before looking at the key; then compare, discuss and fix the rubric.

Prompt 6: fit a rubric to a grading-time budget

Here is a grading rubric (below the line). Grading time is the problem:
we have [N] submissions, [N] graders and about [N] hours in total, so
about [N] minutes per submission. A timed trial with this rubric took
about [N] minutes per submission. Grading is done in [Gradescope / by
hand / other], and an autograder [does / does not] check [what it checks].

Revise the rubric to fit the budget while still assessing every one of
these learning outcomes:
1. [outcome]
2. [outcome]

Use these moves where they fit: fewer and coarser criteria (three levels:
full, partial, none); yes-or-no items for objective checks; moving
mechanical checks to the autograder; one mutually exclusive group of
levels per criterion; feedback written into the items; a cap on how
often each mistake is charged. Do not change the assignment or the
grading policy; list any such changes as suggestions for the instructor
instead.

Give me:
1. The revised rubric as a table.
2. An estimate of the decisions per submission before and after.
3. What feedback students lose, and the one criterion where you would
   keep real written feedback.
4. Suggested changes to the assignment or policy, for the instructor.
---
[rubric]
---

Time yourself on ten submissions with the revised rubric before adopting it.

Is it a good rubric? How to judge the draft

Observable: every criterion names something you can see in the work, not a value word.
One idea per criterion, and the levels are clearly different.
Aligned: each criterion maps to a learning outcome, and each outcome is assessed.
Fair to other correct answers: try the tool's alternative solutions, and one of your own.
Partial credit makes sense: a correct structure with one gap scores more than an answer that goes nowhere.
Fits the course: notation, definitions and expectations match lecture and the instructor, not the tool's defaults.
Tested: you and another grader applied it to real submissions and agreed, or fixed it until you did.
Approved: your instructor has read the final version.