AI at Home
ChatGPT Prompt for Selective Writing Feedback
Ask ChatGPT plainly to mark your child's writing and it gives a mark out of 10, opens with "This is a strong start…" and sometimes rewrites whole paragraphs. The prompt below makes it mark six areas out of 25, quote your child's own sentences as evidence and suggest changes one sentence at a time. We ran both 50 times in total on the model behind ChatGPT. One warning up front: the prompt fixes the feedback, not the mark. On the 8 published NSW samples, ChatGPT's totals came out about 3 marks below the official marks either way.
The Prompt
You are an experienced marker for the NSW Selective High School Placement Test writing task. Mark the piece of writing below. It was written by a Year 5 or 6 student. Task: [paste the task or topic here] Writing: [paste the writing here] Give a mark for each of these six areas (25 in total): - Content & Detail: 0-5 - Structure & Cohesion: 0-5 - Style & Vocabulary: 0-5 - Sentence Variety: 0-4 - Punctuation: 0-3 - Spelling: 0-3 Rules: 1. First, say whether the writing answers the task and uses the form the task asks for. 2. For every area: give the mark, quote one sentence from the writing as evidence, and say why. 3. For every area, pick one sentence to improve and show a better version. Change single sentences only. Do not rewrite a paragraph or the whole piece. 4. Do not praise anything you cannot back with a quote. Do not invent anything that is not in the writing. 5. End with the total out of 25, then the three most important things to fix, in order of importance.
How to use it well:
- Your child writes first, on their own. The real test is typed, 30 minutes, unaided. Practise the same way, then paste.
- Paste the task wording, not just a title. Without it, ChatGPT can't tell whether the piece answers the question or uses the right form (an advice sheet, a diary entry, a story).
- Start a new chat for each piece, so feedback on an earlier essay doesn't colour the next one.
- Read the quoted sentences together, not the total. Pick one or two of the suggested changes, and have your child rewrite those sentences in their own words.
Add One Line for the Type of Writing
| If the task asks for | Add this line to the rules |
|---|---|
| A story (narrative) | 6. Check whether the story has a problem or turning point, and whether the ending connects back to the opening. |
| A persuasive piece | 6. Check whether every main point is backed by a specific example, and whether the other side is answered. |
| An advice sheet | 6. Check whether the advice is practical and specific, written directly to the reader the task names, and easy to follow. |
| A diary entry | 6. Check whether it is written in the first person and shows the writer's thoughts and feelings, not just a list of events. |
| A news report | 6. Check whether the opening tells the reader who, what, when and where, and whether it reports in the third person without the writer's opinion. |
| A letter, email or speech | 6. Check whether the tone suits the person it is written to, from the greeting to the sign-off. |
More on what each form needs is in our guides to narrative, persuasive and informative writing.
What Happened When We Tested It
We used the persuasive opening shown on our homepage: 142 words in two paragraphs, on whether friendship is the most important thing in life. The plain question was "Can you mark my child's writing and tell them how to improve it?". Each version was asked six times, each time in a fresh conversation.
| Plain question | This prompt | |
|---|---|---|
| Mark | Out of 10 every time: 8/10 three times, 7.5/10 three times (two also gave a letter grade, B+ and B) | Six areas out of 25 every time: 14/25 five times, 15/25 once |
| First comment on the writing | Praise, 6 of 6 ("This is a strong start…") | Whether it answers the task, 6 of 6 |
| Rewrote a whole paragraph | 1 of 6 (an 87-word "improved version") | 0 of 6 (longest suggested change: 30 words) |
| Sentences quoted from the writing | 5 to 9 per reply | 16 to 22 per reply |
In the ChatGPT app itself (one run each, same day), the plain question went further: after 7.5/10 and "Your child has made a really good start", it added "a corrected version" of the whole piece. The prompt gave 14/25, with the six marks, a quote for each and single-sentence changes, the same as in our test runs.
Does it actually read the sentences? We deleted all five commas from the same piece and ran the prompt six more times. Punctuation dropped from 2 to 1 in all six replies, and the total to 13 every time (from 14 or 15). Nothing else in the piece changed, so the drop came from the commas.
How we tested: OpenAI's chat-latest model through the API, which OpenAI describes as pointing to the latest model used in ChatGPT; default settings; no custom instructions; tested 1 November 2026; 50 runs in all (18 on the homepage piece, 32 on the NSW samples below). For the homepage piece the task line was "Is friendship the most important thing in life? Write a persuasive piece."; for the NSW samples it said only "Write an advice sheet." The ChatGPT app adds its own instructions and memory, and OpenAI updates the model without notice, so your replies will be worded differently. Check the first reply follows the rules.
Can ChatGPT Mark Selective Writing Accurately?
NSW has published 8 marked sample responses to one Selective writing task (an advice sheet), with official totals from 19 to 25 out of 25. We gave each one to ChatGPT twice in two ways: a plain "Can you mark my child's NSW Selective writing out of 25 and tell them how to improve it?", and the prompt above.
| Plain "out of 25" | This prompt | |
|---|---|---|
| Range of marks given | 18 to 21 | 17 to 21 |
| Average compared with the official mark | 2.9 marks lower | 3.6 marks lower |
| The sample officially marked 25 | 19 and 20 | 20 and 21 |
| The sample officially marked 19 | 18 and 18 | 17 and 17 |
| Marking areas used | Made up its own, and they varied between replies (it marked "Ideas" out of 5, 6, 9 or 10) | The six areas given |
| Rewrote a passage of 50+ words | 12 of 16 replies, including one complete model answer of about 400 words | 0 of 16 (longest suggested change: 35 words) |
Two things stand out. First, ChatGPT squeezes the scale. The official marks for the best and weakest samples are 6 apart (25 and 19); ChatGPT's averages for those two were 1.5 apart with the plain question and 3.5 apart with the prompt. That pattern is common in AI marking, and better wording didn't fix it here: the prompt changed the shape of the feedback far more than the marks. Second, the plain "out of 25" question made up its own marking areas, so its 20/25 isn't measured against the Selective criteria at all.
So read ChatGPT's total as a rough signal, and spend the time on the quoted sentences, which is where the prompt does help. These samples are all strong pieces; we don't have official marks for weaker writing to test against.
For comparison, our own scorer lands within 0.9 marks of the official totals on average on the same 8 samples. In fairness to ChatGPT, our scorer was refined against these samples over many versions, so that is not a blind test. The accuracy page has the method, an independent NSW teacher's check on 23 real essays, and the limits.
Rather not paste a prompt every time? EurekaWrite marks the same six areas out of 25 with a fixed model, quotes the sentences behind every mark, and suggests changes one sentence at a time, never a rewritten essay. With a free account, it also tracks which problems keep coming back. Free to try, no signup.
Score my child's writing →Why the Prompt Is Written This Way
- "Year 5 or 6 student" and the task: asked plainly, ChatGPT had to guess the year level (one reply: "I'm assuming your child is in Years 4–6 (ages 9–11)") and had no task to check the writing against.
- The six areas and their marks: without them, it made up its own areas, and they varied from reply to reply. The 25-mark total and its split into a 15-mark set and a 10-mark set follow the published structure of the test; the six areas inside those are our reading of it. Our marking criteria guide explains each area.
- A quote for every mark: you can check each claim against the writing yourself, and vague praise has nowhere to hide.
- Single sentences only: the child should do the rewriting. An AI-written paragraph teaches very little and can't be used in the test.
- No praise without a quote: this is what stops the "This is a strong start…" opening.
- Three things to fix, in order: a list of fifteen suggestions gets ignored. One clear next step gets done.
Before You Start: Age and Privacy
- Age: OpenAI says ChatGPT is not meant for children under 13, and children aged 13 to 17 need a parent's permission. Most Year 5 and 6 students are 10 to 12. Run the prompt yourself and go through the feedback with your child.
- NSW public schools: students in Years 5 to 12 can use NSWEduChat, the Department of Education's own AI tool, through school. We haven't tested this prompt in it.
- Privacy: don't paste your child's full name, school, or personal details from the writing. Check ChatGPT's data controls if you don't want chats used to train its models.
- Other chatbots: the prompt is plain instructions and can be pasted anywhere, but we have only tested it on ChatGPT's model. Check that the first reply gives six marks, quotes the writing and changes only single sentences.
For the bigger picture on where AI helps with Selective preparation and where it doesn't, see why writing is the Selective component best suited to AI practice.
Frequently Asked Questions
Can ChatGPT mark NSW Selective writing accurately?
It can give useful feedback if you ask in the right way, but treat its mark as rough. On the 8 sample responses NSW has published (official marks 19 to 25 out of 25), the model behind ChatGPT gave 17 to 21, about 3 marks below the official mark on average, with or without a detailed prompt.
Is my child old enough to use ChatGPT for writing practice?
OpenAI says ChatGPT is not meant for children under 13, and children aged 13 to 17 need a parent's permission. Most Year 5 and 6 students are 10 to 12, so a parent should run the prompt and go through the feedback with the child. NSW public school students in Years 5 to 12 can use NSWEduChat, the Department of Education's own AI tool, through school.
Will ChatGPT write the essay for my child?
Asked plainly for a mark and feedback, it often rewrites whole passages: in our test, 12 of 16 replies on the NSW samples included a rewritten passage of 50 words or more, and one included a complete model answer. In the ChatGPT app, the plain question ended with a corrected version of the whole piece. With the prompt on this page, none of 28 replies did; the longest suggested change was 35 words.
Does this prompt work in Gemini, Copilot or other chatbots?
We have only tested it on the model behind ChatGPT. The prompt is plain instructions, so you can paste it into any chatbot, but check that the first reply gives six separate marks, quotes the writing and changes only single sentences.
Practice prompts, a 30-minute timer, and feedback on every piece.