Marking and feedback prompts

Short answer: AI marking belongs in the role of provisional assessment plus specific feedback, with the teacher keeping correction, sampling and decisions. Two sentences belong in every marking prompt — any mark is provisional and requires teacher review, and the student's work must not be rewritten.

Feedback that looked excellent and should not have been issued

A Primary 5 Chinese Language teacher tried AI marking. She pasted one student narrative and asked:

Mark this Primary 5 composition and give a score and comments.

Back came something polished: 18/20, three paragraphs of comment, and a rewritten version of the piece that was noticeably better written than the original.

Three problems at once. First, where did 18 come from? She supplied no rubric, so the model applied a standard it invented, and the next script might be marked far more generously or harshly. Second, the student will copy the rewritten version verbatim. Third, the file she pasted had the student's name and class at the top.

Each of those corresponds to a sentence that should have been in the prompt.

Why this works

Marking is two different jobs: judging the level and saying what comes next. Models are strong at the second, because it is pattern recognition and expression. They are entirely dependent on you for the first — with no rubric, the model marks against an average impression of student writing, which may bear no relation to your school's standard.

An analogy: give a newly qualified teacher your rubric and their marks will land close to yours. Give them nothing and they will mark on instinct, differently on every script. The difference is not effort, it is a shared standard.

So a marking prompt rests on three pillars: supply the rubric, forbid rewriting, label the output as requiring teacher review.

Our article AI marking and instant feedback already covers the three subject scenarios and the three checkpoints teachers must keep. This lesson does not repeat that; it concentrates on the prompts.

Deal with student data first

Before writing any marking prompt:

• Strip identifying data. Names, class, student number, parent names, addresses and internal school IDs — remove them or replace with "Student A".
• Never paste a whole class file into a general tool. A spreadsheet with thirty names and their results is a personal data incident waiting to happen.
• Work on a gated platform. This is precisely what a school platform is for: four-layer gating, teacher monitoring, encryption at rest and an ai_audit_logs trail, none of which a general chat tool has. Our overall approach is on the security page.
• Tell students. They should know AI is the first reader and the teacher is the final marker.

For how safety settings get circumvented and what actually holds, see lesson 14.

Template 1: Chinese composition against a rubric

You are a marking assistant for a Hong Kong primary Chinese Language teacher. Your output is a provisional assessment for teacher review, not a final grade.
Task: provide a provisional assessment and feedback on one student composition against the rubric below.
Context: Primary 5 narrative unit, learning focus is conveying feeling through action and dialogue. The comment will be read by the student, so it must be in language they can understand.
Rubric: content 6 marks (complete event, concrete detail, expressed feeling); structure 5 marks (clear shape, distinct paragraphs); language 5 marks (accurate word choice, varied sentence patterns); punctuation and spelling 4 marks.
Output format: three parts — (1) a four-row table, one per dimension, with the suggested mark, one sentence of justification and a quotation from the student's text; (2) the comment for the student, three to four sentences, the first naming a specific strength with a quotation, the second one area to improve, the third one concrete action to take now; (3) one line headed "for teacher review" listing the two judgements you are least certain about.
Constraints: do not rewrite any of the student's sentences or paragraphs, only indicate a direction; every quotation must be words the student actually wrote; all marks are provisional and the output must open with "provisional — requires teacher review before release"; if the composition contains any personal data, do not reproduce it in your output.

Student composition (de-identified):
[paste]

Template 2: English writing feedback

An English Language teacher marking English writing should prompt in English — the feedback itself is in English and the marking codes must match the school's own.

You are a marking assistant for a Hong Kong secondary English Language teacher. Your output is a provisional assessment for teacher review, not a released grade.
Task: give feedback on one piece of student writing using our marking codes.
Context: Secondary 3, a 250-word argumentative paragraph. The learning focus is topic sentence, supporting detail and concluding sentence. Students are Chinese first language; the most common problems are tense consistency and run-on sentences.
Marking codes: T tense, SV subject-verb agreement, ART article, WW wrong word, RO run-on sentence, ^ missing word, ORG organisation.
Output format: three parts — (1) the student's text reproduced exactly as written, with codes inserted in square brackets at the point of each error and nothing else changed; (2) a table of error counts by code, most frequent first, naming the two the student should prioritise; (3) three sentences of feedback addressed to the student, giving one strength with a quotation, one priority to fix, and one concrete next step.
Constraints: never rewrite the student's sentences or supply a corrected version — insert codes only, so the student does the correcting; do not alter spelling or punctuation in part 1; any marks or bands are provisional and must be labelled as requiring teacher review; do not reproduce the student's name, class or student number.

Student writing (de-identified):
[paste]

Template 3: diagnosing mathematical working

You are a marking assistant for a Hong Kong secondary Mathematics teacher. Your output is a provisional assessment for teacher review.
Task: diagnose a student's working, find the first step that goes wrong, and infer the misconception behind it.
Context: Secondary 2, unit "applications of linear equations". The standard method has five steps: define the unknown, state the relationship, form the equation, solve, check.
Output format: three parts — (1) a three-column table comparing the student's steps with the standard method, marking each correct / incorrect / missing; (2) one sentence naming the first step that goes wrong and the misconception you infer; (3) two practice questions of the same type with different numbers, to test whether the misconception is cleared, with answers.
Constraints: do not rewrite the full solution for the student, identify the failing step and the direction only; do not assume steps the student did not write — mark them missing and note "teacher to confirm with student"; if the student's method is non-standard but mathematically valid, say so explicitly and do not mark it wrong; all judgements are provisional and require teacher review.

Student work (de-identified):
[paste]

Template 4: whole-class error patterns

This is the highest-value prompt in the lesson, because it decides what you teach next.

You are a data analysis assistant for a Hong Kong secondary [subject] teacher. Your output supports a teaching decision.
Task: summarise class performance on one assessment and identify shared error patterns.
Context: [year level], unit "[name]", class of [number]. The learning focus numbers and mark allocation are in the specification table: [paste]. The data I supply is per-item error type counts, with students identified by number rather than name.
Output format: four parts — (1) a table, one row per learning focus, giving class mastery and the two most common errors; (2) three teaching recommendations, each naming which learning focus the next lesson should address first and why; (3) a suggested tiering, estimating how many students need part one, part two and the extension; (4) one sentence identifying which conclusions are limited by sample or item count and should not be over-read.
Constraints: reason only from the data I supplied and do not assume data I did not give; never name individual students, use the numbers; where a learning focus is covered by a single item, state that the conclusion is weakly supported; do not recommend anything requiring extra lesson time.

Data:
[paste]

Common mistakes

• Asking for a mark without a rubric. With no rubric the model's severity drifts between scripts and bears no relation to your school's standard. The rubric is the foundation of the whole prompt.
• Letting AI produce a rewritten version. Students copy it and the feedback becomes pointless. Every prompt touching student work needs the "suggest, do not rewrite" line.
• Treating a provisional mark as final. Sample a portion fully yourself, compare, and keep one-click amendment. Responsibility for any released mark is the teacher's.
• Pasting files that were never de-identified. The easiest mistake to make and the most serious. Strip names and class first, or work on a gated platform.

Going further

• A different subject: for General Studies and Science practical reports, ask for feedback against the four stages of inquiry — question, design, data, conclusion. For Visual Arts appreciation writing, ask it first to identify which visual-element vocabulary the student used, then say whether the student is still describing or has moved into interpretation.
• A different year level: primary comments should be short, in the second person, and contain one action the student can take today. Secondary can include the band and the rubric descriptor so students can see why they are at that level.
• SEN adaptation: for students with writing difficulties, focus feedback on content and reasoning, and add the explicit constraint "do not treat spelling error count as the main judgement; report punctuation and spelling in a separate section".
• Calibrating the register: use the example technique from lesson 3 — paste three comments you wrote yourself and the model adopts your voice instead of an institutional one.

After marking, students hold feedback but do not know how to act on it. That calls for a tutor that gives hints rather than answers. See lesson 10, scaffolded tutor prompts.