Can AI grade essays to my own custom rubric?

What does a rubric need before any tool can apply it?

A rubric a grading tool can apply is the same rubric a substitute teacher could apply without calling you. That is the honest test, and most classroom rubrics fail it on the first try. Four parts do the work: the criteria you are assessing, the performance levels a response can land in, a descriptor of what each level looks like on the page, and the points each criterion carries.

The part usually missing is the descriptors. Criteria get named, levels get labeled emerging through advanced, points get assigned, and then the top band says strong use of evidence while the bottom says weak use of evidence. That is a scale, not a description. It tells a reader which end is better without telling anyone what to look for, which is how two teachers score the same paper two bands apart.

Keep it short. Four to six criteria with sharp descriptors grade more consistently than twelve with overlapping ones. Every extra criterion is another judgment call, and across a stack of ninety essays the vague ones are where the scoring drifts.

• Criteria: the specific things you are assessing, such as thesis, use of evidence, or organization.
• Performance levels: the bands a response can fall into, usually three to five.
• Descriptors: a short statement of what the work looks like at each level.
• Point allocations: how many points each criterion carries, which is also its weighting.

How do you write descriptors a grader can actually check?

Write what is on the page, not how the writing made you feel. “Integrates textual evidence and explains how it supports the claim” is checkable by anyone. “Good analysis” is not. The gap matters for human scoring too, but an experienced reader quietly fills it from memory of what the class was taught, and software has no such memory to draw on.

Keep one idea per criterion. If a criterion covers organization and mechanics together, an essay that is well structured and full of comma splices has no correct level, and whatever you choose is a compromise you cannot explain to the student. Split the criterion, or move one half of it somewhere else.

Then read adjacent levels against each other. If levels three and four could both fairly describe the same paragraph, the boundary is too soft and the scoring will drift whoever applies it. The fastest fix is to name one thing level four does that level three does not, then check that the thing is visible in the text rather than inferred from it.

Test the draft on two or three essays you have already graded. If the descriptors explain why one paper sat above another, they are specific enough. If you are relying on a professional instinct the wording never captures, write down what that instinct noticed and put the sentence in the descriptor. This is worth doing even if no software ever touches the rubric, because it is the same work that makes grade-team calibration go quickly.

How do you tie a custom rubric to Common Core or state standards?

Tag the criterion, not the descriptor. A short reference beside the criterion name is enough: a Common Core writing anchor standard, a state standard such as the Texas Essential Knowledge and Skills or Florida’s B.E.S.T. standards, or the AP scoring guideline you adapted. The descriptor underneath stays in plain language so students can read it without a decoder.

The tag earns its place when a score is questioned. A parent email and an administrator’s file review ask the same thing in different words: how was this number reached? A tagged rubric answers by pointing at the criterion, the descriptor the response met, and the published expectation behind it, instead of asking you to reconstruct a judgment from memory six weeks later.

It also makes the rubric shareable. A grade team that agrees on one version has every section judged against the same criteria, which is the only condition under which section-to-section comparisons mean anything. Hand the same descriptors to students as study guides while they draft, since a well-written descriptor already says what the finished work has to do.

How should weighting and borderline papers be handled?

Point allocation is the weighting. There is no separate dial. If argument carries forty points out of a hundred and mechanics carries ten, that ratio is your public statement about what the assignment values. Check it against what you tell the class matters, because the two often disagree, and students read the point column before they read anything else.

Holistic rubrics work the same way, with one set of bands instead of criterion-by-criterion scoring. Give the band descriptors the same level of detail and a holistic rubric is just as gradable as an analytic one. Analytic scoring produces more specific feedback, holistic scoring is faster to apply, and both hold up if the descriptors are concrete.

Decide the borderline rule in advance. A stated rule, such as scoring down and naming the missing feature in the comment whenever a response sits between two levels, beats deciding case by case at eleven at night. Borderline cases are also diagnostic. If one criterion produces them again and again, the descriptor is the problem, not the papers.

Where do success criteria and comment banks fit?

Three artifacts, three jobs. The rubric is the standard you grade against. Success criteria are the same expectations rewritten for students and handed out before they draft, so nothing in the rubric arrives as a surprise. A comment bank holds the phrases you write over and over, saved once, so the wording stays consistent across a class and across a year.

Build them in that order and only as far as you need. The rubric does the heaviest lifting. Add success criteria when you notice yourself explaining the rubric out loud in the same words every lesson. Start the comment bank by saving the three comments you wrote most often this term, which takes five minutes and pays back on the next set.

How does software apply a rubric you wrote yourself?

This is where a tool helps, and it helps only as far as the rubric allows. A grading tool such as JeddAI takes the rubric you entered, with your criteria, levels, descriptors and point values, and drafts a score and a comment for each response against that framework rather than against a general model of good writing. If your rubric weights argument heavily and mechanics lightly, the drafts follow suit. If a criterion names a specific skill, the comments speak to that skill.

What changes in practice is where your attention goes. The routine pass is applied evenly across the whole class, and your time moves to the papers that need a teacher: the borderline ones, the unusual responses, and the personal note a student will act on. Scores and comments stay drafts until you approve them, so the grade of record is still yours.

Evaluate any tool the way you would test the rubric itself. Enter your own criteria rather than accepting a supplied template, run it over a set of essays you have already graded, and compare the output with the scores you gave. Where you edit heavily, look at the descriptor before you blame the software. Jeddle’s expert essay grading works from the rubric you provide, so that comparison takes one assignment to run.