What should you have ready before your first run?
Three things, and you almost certainly have all of them already. A rubric or achievement standard you genuinely mark against. One class set from a task you were going to mark anyway. And a rough list of the comments you find yourself writing over and over. Setting up an AI marking tool is mostly the work of writing down criteria you have been carrying in your head for years.
Do not start with an internally assessed NCEA piece. Pick something formative and low stakes, where nothing a draft gets wrong can touch a reported result. A practice paragraph, a mid-unit response, a mock question. The first session is for learning the rhythm of drafting and reviewing, not for clearing a backlog.
Block out about forty minutes. Most of that goes into the criteria, and you only do that part once. The marking itself is quick after the setup is done, and the setup carries over to every task that uses the same standard.
Which task should you choose for the first run?
Choose a task with a rubric you trust and a set of responses with a real spread. If everyone landed on roughly the same grade, you learn nothing about how the drafts handle the top and the bottom of the class. A spread shows you quickly whether the suggested comments track your judgement or drift toward a flat, encouraging middle, which is the most common weakness in machine-drafted feedback.
Short extended responses work better than very long ones at the start. Somewhere between a paragraph and a page. Long pieces produce long drafts, and long drafts are hard to check line by line while you are still working out how much to trust them. Once you are confident on short work, scale up to full essays.
- A task you were going to mark anyway, so the time is not additional.
- A rubric or standard you already apply confidently.
- A class set with a genuine range of achievement.
- Formative, low stakes, and nothing that contributes to a reported result.
How do you turn a rubric into criteria a tool can actually use?
Most rubrics are written for a colleague who already knows the subject. Achievement standard descriptors are deliberately holistic, and a phrase like develops ideas convincingly carries a great deal of unstated professional knowledge. A tool has none of that context, so it fills the gap with a guess. The work is to unpack each descriptor into things you could point at in a script.
Split compound descriptors. If one line asks for structure, evidence and control of language, that is three separate checks, not one. Write your success criteria in the words you use with students, because those are the words you want appearing in the feedback. Then write out the ten or fifteen comments you reach for most often, including the ones you use for strong work. That starter bank improves a first draft more than any amount of rewording your instructions.
The discipline here is the same one behind expert essay marking anywhere. Name what is observable in the work rather than what you assume about the writer. If a criterion is too vague for a colleague to apply consistently, it is too vague for a tool, and the drafts will show you that within about five responses.
How do you review a draft without rubber-stamping it?
Read the student’s work first, at least for the first few responses. If you read the draft first you will find yourself agreeing with it, and a review turns into a rubber stamp. Form your own quick impression, then check each drafted comment against the evidence and delete anything the response does not actually support.
Watch for three failure modes. Wording that implies a higher grade than the work has earned. Praise that could be pasted onto any script in the set. And confident claims about a student’s intent that the writing does not demonstrate. Cut all three. Then keep the single strength and the single next step the student most needs, in your own words. A student who receives six comments acts on none of them.
This review step is where your professional judgement stays in charge, and it is also where the time saving is real. Editing a draft that is eighty per cent right is much faster than writing from a blank box, and it is more consistent across a class than marking script twenty when you are tired.
What does NCEA moderation expect from AI-assisted feedback?
The assessment judgement has to be yours. Internally assessed standards are subject to external moderation, and a moderator is checking whether your grade and the evidence in the student’s work line up against the standard. None of that changes because a draft helped you write the comment. What matters is that the comment points at evidence and uses the language of the standard rather than a generic quality scale.
Two habits keep you safe. Keep feedback anchored to the descriptor it relates to, so the trail from evidence to judgement stays visible to anyone who looks at the file later. And check your school’s position on student data before you upload anything identifiable to any system. Most schools now have a policy, and asking first is easier than unwinding it afterwards.
It also helps to be explicit with students about which activity is happening. Formative feedback on a draft is a different thing from assessing final work against a standard, and saying so out loud makes the conversation about their own use of AI a lot simpler. Pair the feedback with study guides or exemplars for the standard so the next step has somewhere to go.
What does a manageable first session actually look like?
Prepare the criteria once. Run one piece of work all the way through. Review the draft against the evidence. Then note what you changed. That last step is the one people skip, and it is the one that improves the next run, because the edits you keep repeating are telling you exactly where your criteria are vague.
Expect the first piece to take longer than marking it by hand. The second is faster. By the fifth or sixth you have a reasonable sense of where the drafts hold up and where they need work, and that is the point at which running a full class set is sensible. Give it two or three sessions before deciding whether it earns a permanent place in your routine.
- Set up one rubric, its success criteria, and a small comment bank.
- Run a single piece of student work end to end.
- Read the work, then check each drafted comment against the evidence.
- Cut unsupported claims and generic praise, keep one strength and one next step.
- Note which edits you made twice, and fix the criterion that caused them.
How does a purpose-built marking tool fit into this?
Everything above works with any drafting tool. What a purpose-built marking platform changes is where the criteria live. In a general chatbot you paste the rubric in again every time, and consistency across a class depends on you pasting the same thing thirty times. In a platform built for marking, the rubric, the success criteria and the comment banks are stored once and applied to every response in the set.
Jeddle is built around that loop for secondary teachers. Its feedback engine, JeddAI, drafts comment by comment against the criteria you have entered rather than against a general model of good writing, and it reuses your own comment bank phrases so a class set reads consistently. Nothing reaches a student until you approve it.
In New Zealand the practical payoff is the moderation trail. Because JeddAI drafts against the descriptors you entered, the comments a student sees already use the language of the standard, which is what a moderator is looking for. You still make every judgement about grade and wording. The tool does the typing, not the marking.
- Rubric, success criteria and comment banks stored once, applied to a whole set.
- Drafting anchored to the descriptor each comment relates to.
- A review step where you edit or delete before anything is released.
- Feedback that keeps the language of the achievement standard for moderation.
