You have 38 Class 9 answer sheets for a four-mark question, a free Sunday and a nagging feeling that the answer you marked as 3/4 at 9 am would have been 2/4 by 4 pm. That drift is the real marking problem. AI prompts for teachers can help with it, but only if you use the model for what it is decent at: drafting a clear rubric, flagging inconsistency, and turning your notes into readable feedback. The mark stays yours.
This post is about assessment, rubrics and feedback. It doesn't repeat lesson planning or worksheet design, which are covered in AI lesson plans and worksheets for CBSE teachers and the wider prompt guide for teachers in India. These are untested templates. I haven't run them on a class set, and the worked example below is illustrative. Adjust to your board's marking scheme and your school's policy.
Rule zero: no student personal data in public tools
Before any prompt, decide what goes in the box.
- Never paste names, admission numbers, photos of students, parent details, health or counselling notes, or anything that identifies a child.
- Replace names with codes (S01, S02) and keep the code list in your own register, offline.
- Paste the answer text only. If the answer itself contains a name or an address, remove it.
- Use only tools your school permits. A free personal account is a different data deal from a school-approved one, and the terms change. Read them for the exact plan you use. The what not to paste into ChatGPT guide covers the general version.
- Rules about children's data differ by country and by board. I haven't stated any here. Ask your school or a qualified adviser.
If the rule makes a task impractical, do the task by hand. A rubric needs no student data at all, which makes it the best place to start.
How do I build an assessment rubric with AI?
A rubric is a set of observable differences between a 4, a 3, a 2 and a 1. Models are good at drafting these because they have seen thousands of rubrics. They are bad at knowing what your class actually needs, so give them the learning outcome, the question and your constraints.
You are helping a [SUBJECT] teacher design a marking rubric.
Learning outcome being assessed: [OUTCOME, copied from the syllabus]
Question or task: [PASTE THE QUESTION]
Marks available: [N]
Class level: [CLASS / AGE], mixed ability
Board / marking style: [e.g. step marking, holistic]
Produce a rubric table with 3 to 4 criteria. For each criterion give
descriptors for every mark band. Rules:
- Descriptors must be observable in the written answer ("states the formula
and substitutes values"), never vague ("good understanding").
- Each band must differ from the next by one clear, checkable thing.
- Do not reward length, handwriting, or fluent English unless that is the
outcome being assessed.
- List what a student could do that would earn marks but that I have not
anticipated, as questions for me to decide.
Finish with a one-line "common wrong answers to expect" list, labelled as
guesses I must check against real scripts.
Two lines do the heavy lifting. "Observable in the written answer" removes mush like "demonstrates insight". "Do not reward length" counters a known habit of models and humans alike: longer answers look better. Then read the table with a pencil. Delete any descriptor you could not defend to a parent who asks why their child got 2.
Check the rubric against your own wrong answers
Models invent plausible student errors. Yours are better. After the draft, add three real anonymised wrong answers from last year and ask:
Here are three anonymised student answers to the question above (S1, S2, S3).
Without giving final marks, list for each: which rubric descriptors it
clearly meets, which are unclear, and where the rubric wording leaves me
unsure how to mark it. Suggest rewording only for the unclear descriptors.
The point is the "unclear" list. Every ambiguity found now is an argument you won't have in a re-evaluation request.
Can AI help me mark more consistently?
Use it as a second reader on a small sample, not as the marker. A sensible pattern:
- You mark 8 to 10 scripts by hand using the rubric.
- You paste those answers (codes only) and your marks into a prompt and ask where the rubric and your marks disagree.
- You decide whether the rubric or your marking drifted, and fix that.
Rubric: [PASTE]
Below are [N] anonymised answers with the mark I gave and my one-line reason.
For each, say whether my reason is consistent with the rubric wording.
Flag pairs where a similar answer got different marks, or a weaker answer
got a higher mark. Do not give your own marks. Do not assume I am wrong.
Quote the answer text that supports each flag.
[S01 | my mark | my reason | answer text]
...
"Do not give your own marks" matters. If you ask for a score, you invite a confident number you may anchor on. A flag that says "S03 and S07 both skip substitution, but S03 got 3 and S07 got 2" is something you can resolve in a minute.
Be honest about the limits. Models vary between runs, so asking twice can give different flags. They can be swayed by polish and length. They cannot read poor handwriting reliably, and scanning student scripts into a public tool is a separate privacy decision. For high-stakes or board-linked assessment, follow your board's rules about marking, which I haven't researched here.
How do I turn my notes into student feedback?
The best feedback names one thing done well, one thing to fix and one next step, in language the student can act on. Models can write that if you supply the substance. The failure mode is the opposite: you paste an answer and ask for "feedback", and you get a warm, vague paragraph that could apply to anyone.
Supply a note per student. Three words are enough.
Write feedback for a [CLASS] student on a [TASK] using this rubric:
[PASTE RUBRIC SUMMARY]
My observation about this student's work (use this and nothing else):
- Strong: [e.g. clear labelled diagram]
- Fix: [e.g. units missing in the final line]
- Next step: [e.g. redo question 3 and check units before submitting]
Rules:
- Max 3 sentences, plain English, [Hindi/English/mixed] as I specify: [LANG].
- Address the student as "you". Do not mention marks unless I give them.
- No stock openers ("Great effort", "Well done on your work").
- Do not add praise or criticism that is not in my observation.
- Do not guess the student's effort, attitude or home situation.
Then the discipline: read every output before it goes anywhere, and swap in the real name only at the end, on your own device. A comment that says "you clearly worked hard" about a child who didn't is worse than none. If you teach in a language the model handles unevenly, run the translated version past a colleague or use the checks in translating with AI into regional languages.
Batch tip: keep your observations, not the AI text
Keep a simple sheet with columns Code, Strong, Fix, Next step. The sheet is your record and survives a change of tool. The generated sentence is disposable. If a parent asks what was said, you can reproduce it from your notes.
How do I check feedback and rubrics for bias?
Bias in marking is a fairness issue, not a technical one, and a model can reproduce patterns in its training data. You can test for it cheaply.
Swap test. Take one anonymised answer. Make two versions that differ only in something that should not matter: a regional spelling, an informal phrase, a mention of a gendered pronoun. Run the same feedback prompt on both. If the tone, the praise or the score shifts, the combination of prompt and model is leaking. For a runnable version of this idea, see the counterfactual names bias test script.
Format test. Take a correct answer written in bullet points and the same content in a paragraph. A rubric that rewards one format when the outcome doesn't need it is a hidden bias.
Language test. Ask yourself whether the rubric is scoring the science or the English. If you are assessing explanation of photosynthesis, a correct answer in simple English should score the same as an ornate one.
Pattern check. At the end of a term, look at your own marks by group (section, gender, home language) in a spreadsheet. A model won't do this for you honestly, and you shouldn't paste the class list to find out. A gap isn't proof of bias, but it is a prompt to look at scripts again.
An illustrative example
Invented scenario to show the shape of the output, not a real run. Question: "Explain why the sky appears blue." Four marks, two criteria.
Criterion 1: Scientific explanation (0-2)
2: Names scattering of sunlight by air molecules and states shorter
wavelengths scatter more.
1: Mentions scattering but does not link it to wavelength.
0: Gives a non-scientific reason (for example, reflection of the sea).
Criterion 2: Completeness of reasoning (0-2)
2: Connects the scattering to what the observer sees at midday.
1: States the conclusion without the link.
0: No link to observation.
Questions for the teacher: Does "Rayleigh scattering" need to be named, or
does the description suffice? Is "reflection of the sea" worth partial credit?
Notice the closing questions. They are the useful bit: the model has surfaced two decisions only you can make.
Where this goes wrong
| Failure | What it looks like | What to do |
|---|---|---|
| Vague descriptors | "Shows good understanding" | Require observable wording; delete what you can't defend |
| Model as marker | You anchor on its number | Ask for flags and quotes, never scores |
| Generic feedback | Same warm sentences for everyone | One teacher observation per student; ban stock openers |
| Invented praise | Comments about effort you didn't see | "Use this and nothing else" rule; read before sending |
| Student data leaks | Names or photos pasted in | Codes only; approved tools only |
| Biased tone | Different warmth for different styles of writing | Swap test; format test |
| Run-to-run drift | Different flags each time | Treat as a reader's note, decide yourself |
What not to automate
- The final mark. Judgement about a child's work, including borderline cases, is yours.
- Report-card remarks about behaviour, attendance or wellbeing. These need what you know about the child, and the data is sensitive.
- Anything involving a child's identity or circumstances. Counselling, special-needs and disciplinary notes stay out of public tools.
- Appeals and re-evaluation replies. Answer these from your own records and the board's process.
- Unchecked output. If you wouldn't be comfortable defending the sentence to the student's parent, don't send it.
For the broader ideas behind these prompts, giving context explains why the inputs matter more than the instruction, avoiding hallucinations explains the "use only this" pattern, and the prompt library has more templates to copy.
