How to Build a Rubric with AI—and Keep Every Criterion Measurable

AI can reduce the time it takes to produce a first rubric draft, especially when you are adapting an existing assignment for a new year group, subject, or format. But a fast rubric is not automatically a fair or useful one. The quality of a rubric depends on whether its criteria describe observable evidence, whether its performance levels are distinct, and whether teachers can apply it consistently.

This is where a rubric generator can be helpful: use it to propose structure, wording alternatives, level descriptors, and draft weights. Then apply professional judgement to decide what learners should demonstrate, what evidence counts, and how much each element matters. AI can support the drafting process; the teacher, tutor, or course designer remains responsible for the assessment decision.

Start with the assessable outcome, not a list of generic skills

Before writing any criterion, identify what the assignment is intended to reveal. A report, presentation, practical demonstration, essay, video, or project may involve many skills. A rubric should not attempt to score every possible skill simply because it appears in the work.

Write one short outcome statement that completes this sentence: “By completing this task, learners will show that they can…” For example:

• …make a claim and support it with relevant evidence.
• …explain a scientific process using accurate subject vocabulary.
• …create a proposal that responds to a stated client brief.
• …deliver a short presentation for a defined audience and purpose.

This statement becomes a filter. If a proposed criterion does not help you judge that outcome, it may not belong in the rubric. This protects against a common problem: adding broad items such as “effort,” “creativity,” or “professionalism” when they have not been defined or taught as part of the task.

Design criteria around evidence that can be seen or heard

A measurable criterion tells the learner what evidence the assessor will look for. It does not require a criterion to be entirely numerical. Rather, it should point to features of the work that an assessor can identify and justify.

Compare these examples:

The stronger versions still require teacher judgement, but they make that judgement more transparent. They identify an action, a product feature, or a task requirement.

Use a simple criterion formula

A useful drafting formula is:

Action or quality + content or evidence + task context

For example, “Develops a clear argument using relevant evidence from the provided materials.” This is usually more useful than “Argument” because it indicates what the learner must do and what the assessor should look for.

When prompting a rubric generator, provide the assignment brief, learner level, subject, expected output, and the learning outcome. Ask it to produce no more than four to six draft criteria. Too many criteria can make marking slow, dilute attention, and create the impression of precision without improving the assessment.

Separate criteria that assess different things

Each criterion should ideally assess one main dimension of quality. If one row combines several unrelated elements, a learner may be strong in one element and weak in another, making the score difficult to explain.

For instance, “Research, structure, grammar, and presentation” is likely too broad. It combines content selection, organisation, technical accuracy, and visual or formatting choices. Better options are to separate these into criteria where each matters, or to remove lower-priority items from the scored rubric and give them as formative feedback instead.

Ask three diagnostic questions of every criterion:

1. Can I point to evidence in the learner’s work?
2. Could two learners perform differently on this criterion even if their overall work is similar?
3. Can I explain why one performance level fits better than another?

If the answer is no, revise the wording, split the criterion, or decide whether it should be assessed at all.

Choose weights that reflect instructional priorities

Weights tell learners what matters most. They should therefore be driven by the assignment purpose, not by a desire to make every row look equal.

If an assignment is chiefly about reasoning with evidence, that criterion may deserve more weight than layout or surface-level accuracy. If learners are completing a practical task where safety procedures or technical process are central, those elements may need greater emphasis. The key is alignment: the weighted score should communicate the learning priorities you have taught.

A straightforward approach is to allocate points out of 100, although smaller totals can work equally well. Begin by ranking criteria from most to least important. Then distribute points proportionally. For a short evidence-based written response, a draft might look like this:

These are illustrative weights, not a universal template. A different task may justify a different distribution. Avoid assigning a heavy weight to an easy-to-check feature merely because it is convenient to mark. Convenience is not the same as importance.

Write performance levels that show a real progression

Performance levels describe what changes as work becomes more successful. Labels such as “Excellent,” “Good,” “Satisfactory,” and “Needs improvement” may be useful shorthand, but the descriptors do the real work.

A meaningful progression usually changes one or more of the following:

• Accuracy: from incorrect or inconsistent to accurate and controlled.
• Completeness: from partial coverage to sustained coverage of the task.
• Relevance: from loosely connected choices to purposeful, well-selected choices.
• Reasoning: from assertion to explanation, analysis, or justified decisions.
• Independence: from substantial support to confident, appropriate decision-making.
• Control: from uneven execution to consistent use of a skill or convention.

Do not simply repeat the same descriptor with different adjectives. “Uses evidence well,” “uses evidence very well,” and “uses evidence exceptionally well” leave too much room for individual interpretation. Instead, describe the difference in the evidence itself.

For example, a four-level criterion for evidence might progress from selecting little or irrelevant evidence, to selecting some relevant evidence with limited explanation, to selecting relevant evidence and explaining its connection to the claim, to selecting well-chosen evidence and integrating a clear explanation of how it strengthens the reasoning.

Keep level language proportionate

High-performance descriptors should not require unnecessary complexity, length, or polish unless those features are genuinely part of the outcome. Similarly, the lowest level should describe the work that is present, rather than becoming a vague statement about learner ability or motivation.

Use neutral, work-focused language. Describe the response, not the person. This keeps feedback actionable: learners can see what to retain, revise, or add next time.

Use AI prompts that request evidence, not inflated language

AI often produces polished-sounding criteria that need tightening. A more specific prompt can improve the draft. Include the task details and ask for constraints.

Create a four-criterion analytic rubric for this assignment. Each criterion must describe observable evidence in learner work. Use four performance levels with clear differences in accuracy, completeness, reasoning, or control. Avoid vague words such as “good,” “excellent,” “effective,” and “creative” unless they are defined by evidence. Suggest weights that total 100, but explain the instructional rationale for each weight.

Then review the output carefully. Check that the language matches what learners have actually been taught, that the rubric does not introduce hidden expectations, and that its reading level is accessible to its users. A rubric should clarify the task, not become another difficult text to decode.

Calibrate before the rubric affects grades or high-stakes decisions

Teacher calibration is the process of checking whether assessors interpret and apply a rubric similarly. It is especially valuable when more than one person marks work, when a programme uses common assessments, or when the rubric will inform consequential decisions.

A practical calibration routine can be simple:

1. Select two or three anonymised sample responses representing different levels of quality.
2. Ask each assessor to score the samples independently using the draft rubric.
3. Compare scores and, more importantly, the evidence each assessor used.
4. Discuss where descriptors allowed different interpretations.
5. Revise ambiguous wording, overlaps between levels, or unrealistic expectations.
6. Repeat with another sample if major differences remain.

Calibration is not about forcing identical professional judgement in every borderline case. It is about making the basis for judgement more shared, explainable, and consistent. Keep a short record of agreed interpretations, particularly for examples that often create disagreement.

Build in a learner-facing quality check

Before submitting, learners should be able to use the rubric to review their own work. This is more likely when the rubric is concise and written in accessible language. Consider adding a brief checklist alongside the scored version:

• Have I answered the task and addressed the intended audience?
• Have I included the required evidence, examples, or steps?
• Have I explained my choices rather than only listing them?
• Have I checked the task-specific conventions named in the brief?

This does not replace teacher feedback. It gives learners a clearer route to act on the expectations before assessment.

Use a rubric generator as a drafting partner, then make it yours

A strong rubric is an instructional tool as well as a marking tool. It reveals what quality looks like, supports focused feedback, and can make assessment conversations more concrete. AI can help you begin quickly, generate alternatives, and adapt a structure across tasks. It cannot determine the educational priorities of your class or programme without your review.

If you want a structured starting point, try SubSchool’s AI Rubric Generator to draft criteria, performance levels, and possible weightings. Treat the result as a working draft: check every criterion against the assignment, test it with sample work, and retain the final decision as the teacher or course designer.