Rubric Generator Prompt for Teachers

What this prompt does

A rubric generator prompt turns an assignment brief into a scoring grid: the criteria you'll judge, a description of each performance level, and the wording that separates one level from the next. Carnegie Mellon's Eberly Center defines a rubric as a scoring tool that explicitly describes the instructor's performance expectations, built from criteria, descriptors and levels of performance. This prompt writes all three from the assignment you already set.

The output it's built around isn't the grid. It's the paragraph after the grid, where the model names the places two graders will disagree — because that's the part you can't get from a template.

The prompt

Write a grading rubric for the assignment below. Follow every rule.

**The assignment as students receive it:** [PASTE THE FULL BRIEF]
**Subject and level:** [e.g. "10th grade English" / "second-year undergraduate economics"]
**What this assignment is actually meant to assess:** [THE LEARNING OUTCOME, IN YOUR WORDS]
**Total points and levels:** [e.g. "40 points, 4 levels" / "pass-fail with a distinction band"]
**Where students went wrong last time:** [COMMON FAILURE MODES, or "first time running this"]

Rules:

1. Four to six criteria. Not more. A rubric with nine criteria gets skimmed, and the ninth is always the one nobody scores the same way twice.
2. Every criterion must name observable evidence in the work itself. Never "quality of argument" — write what a strong argument does on the page.
3. Describe each level by what IS there, not by comparison. Banned phrasings: "less developed than", "somewhat", "adequate", "shows good understanding". If a descriptor would fit three different essays at three different grades, rewrite it.
4. No criterion that scores effort, attitude, participation or neatness unless I named it as a learning outcome above.
5. Weight the criteria so the outcome I named carries the most points, and say in one line why each weight is what it is.
6. Match the level count I asked for exactly. Do not add a level to make the maths tidier.

Return three sections:

**A. The rubric.** A markdown table: criteria as rows, performance levels as columns, points in each cell. Descriptors in full sentences, not fragments.

**B. Split decisions.** For each criterion, one sentence naming the single test that separates each pair of adjacent levels. Write the test so I could apply it to a paper at 11pm and get the same answer I'd get at 9am.

**C. What this rubric doesn't measure.** Anything the assignment asks for that this grid can't score honestly. Do not invent a criterion to cover it — tell me it's uncovered and why.

Use only my assignment brief. Don't assume standards, curricula or grade boundaries I haven't given you.

How to use it

1. Paste the brief students actually see, not your summary of it. The gap between the two is where most rubric mismatches come from — you grade the assignment you meant to set, students wrote the one on the handout.
2. Read section B first. The split decisions are the rubric's real content. If the test separating a 7 from an 8 is "somewhat stronger analysis", the criterion isn't finished and you'll feel it on paper thirty.
3. Score two papers you've already graded. Take one strong, one middling, from last year's pile. If the rubric lands within a grade band of what you gave, it's usable. If it doesn't, the disagreement tells you which criterion is mis-weighted.
4. Cut a criterion rather than adding one. Every criterion you keep costs attention on every paper you mark.
5. Give students the rubric with the assignment. A rubric handed out only with the grade is a justification, not an instrument.

Faculty-developed VALUE rubrics are now downloaded by educators in 160+ countries and used for assessment at 3,000+ colleges and universities, per AAC&U's own figures. Worth reading before you write your own from scratch — for the common outcomes like written communication or critical thinking, adapting one of the 16 beats starting from a blank page.

Which rubric type to ask for

Name the type in the prompt. Left unspecified, all three models default to analytic with four levels, which is the right guess often enough that it hides the choice from you.

Example output (partial)

Assignment: 1,200-word argumentative essay on renewable energy policy. Level: 10th grade English. Total: 40 points, 4 levels. Outcome named: using evidence to support a contestable claim.

A. The rubric (2 of 4 criteria shown)

B. Split decisions

• Claim, 4 vs 3: Read only the topic sentences. If one of them could belong to a different essay, it's a 3.
• Claim, 3 vs 2: Ask whether a classmate could write the opposite essay. If not, the thesis is a topic, and it's a 2.
• Evidence, 4 vs 3: Count unsourced figures. Zero is a 4, one or two is a 3.

C. What this rubric doesn't measure

Nothing here scores whether the student's position is correct, and it shouldn't — the outcome you named is using evidence to support a contestable claim, and a well-evidenced essay you disagree with meets it. The brief also asks for "a clear voice", which I've left uncovered: I can't write a descriptor for voice that two graders would apply the same way, and a criterion nobody scores alike is worse than an absent one.

Section C is doing the work there. A rubric that quietly grows a "voice" criterion is how a grade dispute starts, and the honest version puts the gap in front of you before a parent does.

Variations

For a group project

"Add a criterion that scores the individual contribution log, not the group artifact. Describe what evidence of contribution looks like in the log itself."

For a lab report or practical

"Order the criteria in the sequence a student performs the task. Where a step's failure invalidates later steps, say so in the descriptor rather than double-penalising across criteria."

To convert an existing rubric

"Here's the rubric I use now. Rewrite only the descriptors that compare levels to each other, and leave my criteria and weights untouched. List every descriptor you changed and why."

For a single-point rubric

"Write only the proficient column. For each criterion, add two prompts I can use when annotating: one for work above proficient, one for below."

Common pitfalls

• Don't accept adjective ladders. "Excellent / good / fair / poor" is four synonyms wearing a table's clothing. If the model returns one, paste rule 3 back at it and ask again.
• Don't let the points drive the criteria. Asked for 100 points, models invent a fifth criterion to reach a round number. Name the criteria count and the total separately.
• Don't feed it student work. Public AI tools are the wrong place for anything covered by FERPA, and the rubric doesn't need it — the assignment brief is enough.
• Don't ship it unmarked. A rubric you haven't tested on two real papers is a hypothesis.
• Try instead: write the "secure" column yourself, then ask for the levels below it. You know what excellent looks like in your subject; describing the middle is the part that's genuinely tedious.

ChatGPT and Claude both keep the table structure across long rubrics. Gemini tends to compress descriptors into fragments, so add "descriptors must be full sentences" if you're pasting straight into a handout. All three drift back toward comparative wording as the table gets wider — rule 3 is the one worth restating in a follow-up message.

Who uses this prompt

• Teachers: turning a unit assignment into something markable by Sunday night, and handing students the criteria up front. The full set is in prompts for teachers.
• College professors: norming a rubric across TAs, where section B is what actually gets shared at the calibration meeting. More in prompts for college professors.

For the student side of the same table — writing to a rubric rather than writing one — the essay outline prompt takes "what the rubric rewards most" as an input and builds the outline around it.

Both OpenAI and Anthropic make the same point about specificity, and rubrics are an unusually literal case of it: the prompt's ban on the word "adequate" is the entire difference between a grid you can mark with and a grid you argue with.