What prompt turns three observations into a report card comment?
The prompt is a grade band, a subject, three observations, a tone and a length, with the student's name left out. We ran this exact text on 27 September 2026: "You are helping a teacher draft report card comments. Use only the notes below; do not invent achievements, grades or behaviour I did not list. Grade band: Grade 4. Subject: Mathematics. Observations: 1. Solves multi-step word problems accurately. 2. Rushes written explanations and often skips showing the steps. 3. Helps table partners without giving the answer away. Tone: warm and direct, written to the parent. Write three different comments of 50 to 70 words each. Refer to the student as 'your child', never a name. Each comment states one strength with evidence, one area to grow, and one specific thing to practise at home. Number them 1 to 3. No headings, no bullet points."
Three models got the same prompt through OpenRouter, one turn each, default settings. All nine comments kept to "your child", stayed inside the word cap and invented nothing that was not in the notes.
| Model (27 September 2026) | Time to answer | Cost for three comments | Words per comment | Reasoning tokens spent |
|---|---|---|---|---|
| GPT-5.6 Luna | 4.6 s | $0.0008 | 60, 61, 57 | 386 |
| Claude Sonnet 5 | 7.4 s | $0.0061 | 60, 60, 59 | 258 |
| Gemini 3.7 Flash | 12.3 s | $0.0080 | 56, 56, 60 | 1,898 |
Comment 1 from each, unedited. GPT-5.6 Luna: "Your child demonstrates strong mathematical thinking by solving multi-step word problems accurately and supports table partners thoughtfully without giving away the answer. A next step is slowing down written explanations and showing each step clearly. At home, ask your child to solve one multi-step problem and write a complete explanation, including the operation used and how they checked the answer."
Claude Sonnet 5: "Your child solves multi-step word problems accurately, showing strong number sense and problem-solving skills. However, written explanations are often rushed, with steps skipped along the way. At home, please encourage your child to slow down and write out each step of their thinking, even when the final answer is already correct, to build stronger habits for more complex work ahead."
Gemini 3.7 Flash: "Your child demonstrates strong math skills by accurately solving multi-step word problems. However, your child tends to rush through written explanations and often skips showing the required steps. At home, you can support your child's learning by having them practise writing out each calculation step clearly and explaining their thinking for multi-step problems before moving on."
The edit each one needs is the same: cut "demonstrates" to "shows", drop the second "however", and add the one detail the notes did not carry, which is the thing only you know. Gemini's bill is the odd number in the table because it spent 1,898 tokens thinking before it wrote 172 words; for a whole class of 30, the same run costs about $0.02 on GPT-5.6 Luna, $0.18 on Claude Sonnet 5 and $0.24 on Gemini 3.7 Flash at September 2026 OpenRouter rates.
Should I put student names in an AI report card prompt?
No. A comment drafted from "Grade 4, three observations" is not about anyone; the same text with a name, a student number or a diagnosis in it carries what 34 CFR 99.3 lists as personally identifiable information from an education record, and sharing it outside the school is the question FERPA governs. The US regulation under FERPA (34 CFR 99.3) lists personally identifiable information starting with "The student's name; the name of the student's parent or other family members" and running through student numbers, biometric records and date of birth. Among the page's consent exceptions, the one that covers an AI vendor is a party "to whom an agency or institution has outsourced institutional services or functions" under the school's direct control, which is the contract your district signs with an approved tool and the contract you do not have with a consumer chatbot.
The UK guidance is blunter. The Department for Education's generative AI in education page, updated 12 August 2025, says "It is recommended that personal data is not used in generative AI tools" and warns that many free tools use inputs to train their models. That warning is literal for consumer ChatGPT: as of September 2026 OpenAI's help center says Business, Enterprise, Edu and API traffic is excluded from training by default, and a Free, Plus or Pro user has to turn off "Improve the model for everyone" under Data controls to get the same treatment.
We read all nine tool pages on 27 September 2026. Varsity Tutors and Knowt mark the student name as a required field, Monsha takes a CSV or Excel class list with names and pronouns, and ReportCards.ai imports students from Google Classroom. None of the nine mentions FERPA in its page text; several link to a privacy policy, and MagicSchool and Khanmigo sell district contracts that would put them inside the school official exception. Whether your school has signed one is a question for the person who runs student data at your district, and the answer is on file, so ask before you paste. Until then the working rule is the one in what never to paste into an AI chat: the prompt carries observations, the report card carries the name, and you join them after.
Is it OK for a teacher to use AI to write report card comments?
Yes, where the district allows the tool, and the direction of policy is toward allowing it with conditions rather than banning it. Kearny Schools in New Jersey told NBC New York on 31 August 2026 that teachers get access to three LLM platforms for worksheets, quizzes and report card comments, with the superintendent, Flora Encarnacao, calling it a tool and saying it does not replace the teacher's own feedback or the relationship with the student.
The unions have written the conditions. The National Education Association's sample school board policy on AI, dated 23 June 2025, says a district "must inform educators, parents, and students of what and how AI tools are used", and its 2024 policy statement keeps educators and students at the center of every decision the tool touches. The Elementary Teachers' Federation of Ontario's member advice of 15 November 2024 says to "only use digital tools and resources that are approved" and to be aware of privacy when typing anything into an AI tool. Ontario matters here because "ai report card comment generator ontario" is its own autocomplete suggestion; the province's learning skills comments are a known grind.
The disclosure question is older than the policies. At Georgetown Day School in Washington, DC, the student paper reported in March 2023 that teachers had used ChatGPT for first-semester comments; one teacher said the tool helped with a load of more than 50 comments, and one senior called it "freaky and a little bit of a betrayal of trust". The parent objection in that piece was not the tool, it was the idea that attention was being saved at five to ten minutes a child. A comment built from your own three observations and edited by you answers that objection; a comment generated from a grade and a name does not.
The scale of use is now measured. Gallup's survey of 2,232 US public school teachers, fielded 18 March to 11 April 2025 and published June 2025, found 60% used AI for their work that year, 32% at least weekly, and weekly users saving 5.9 hours a week, which Gallup puts at "the equivalent of six weeks per school year". So the decision rule: check whether your district names an approved tool (use it if so), keep names and grades out of anything it has not approved, and tell a parent who asks that you drafted with AI from your own notes and edited every line. That last sentence is only true if you did.
19 example report card comments, AI-drafted and edited
The three comments shown from the run above plus the sixteen below make 19. We drafted the sixteen on GPT-5.6 Luna on 27 September 2026 with the same prompt shape (four grade bands, four situations, 40 to 55 words, a [Name] placeholder, "they" as the pronoun) and then edited every one: "diligently", "imaginative" and "demonstrates" went, spelling went to US, and each closing sentence became one thing a parent can do this week. Swap the placeholder for the name in your gradebook, never in the chat.
| Band | Situation | Comment (drafted 27 September 2026 on GPT-5.6 Luna, edited) |
|---|---|---|
| K-2 | Strong across the board | [Name] comes to every lesson curious and ready. They follow directions the first time, share ideas that move the group along, and finish tasks on their own. The next step is explaining their thinking in more detail. At home, ask them to describe how they solved a problem in full sentences. |
| K-2 | Strong, one clear gap | [Name] is building strong reading habits: they recognize familiar words, retell the main events and choose books with real interest. Number facts are still shaky, and they find it hard to explain a math strategy. At home, play short counting games and ask for two ways to solve a simple addition. |
| K-2 | High effort, skill lagging | [Name] keeps trying when a task is hard, which is the habit that matters most at this age. Letter formation is still developing, so their ideas can be hard to read. At home, five relaxed minutes of printing a day, watching starting points and spacing, will help. |
| K-2 | Behaviour is the concern | [Name] shows real creativity in hands-on activities and contributes well once they settle. They are working on listening, waiting for a turn and following a direction without a reminder. At home, practice one routine: listen, repeat the direction back, finish one step, then choose the next activity. |
| 3-5 | Strong across the board | [Name] understands new material quickly and uses class time well. They back up ideas with details, apply strategies accurately and work respectfully with partners. The next step is revising their own work before handing it in. At home, ask them to reread one assignment and find one change that makes it clearer. |
| 3-5 | Strong, one clear gap | [Name] writes clearly, with organized paragraphs and relevant examples. Multiplication and division facts are not yet automatic, which slows them down on multi-step problems. At home, spend a few minutes a day on facts with cards or a game, then ask how one fact helps with a related problem. |
| 3-5 | High effort, skill lagging | [Name] works hard and takes feedback well, especially on the assignments that stretch them. Reading fluency is still developing, so longer texts take extra time. At home, ten minutes of reading aloud a day, rereading one short passage and talking about what it meant, is the most useful practice. |
| 3-5 | Behaviour is the concern | [Name] has good ideas and contributes well when they are on task. They are working on staying with an activity, using materials responsibly and letting others work without interruption. At home, set a short homework timer, clear distractions, and afterwards ask which one choice helped them focus. |
| 6-8 | Strong across the board | [Name] understands the course material and applies it thoughtfully. They use evidence in discussion, keep assignments organized and act on feedback. The next step is considering more than one side of a question. At home, ask them to explain an issue with an example and then argue the opposite view. |
| 6-8 | Strong, one clear gap | [Name] reasons well in science, using observations and evidence to explain what happened and why. Organization is the gap: missing materials and late work have cost marks this term. At home, a weekly planning routine listing tasks, materials and a realistic finish time would remove most of that. |
| 6-8 | High effort, skill lagging | [Name] puts in steady effort and asks for help when unsure, which is exactly right. Written analysis is still developing; explanations need stronger evidence and a clearer link to the main idea. At home, ask them to write a paragraph as claim, evidence, explanation and one linking sentence. |
| 6-8 | Behaviour is the concern | [Name] produces thoughtful project work when fully engaged. Consistent participation, respectful listening and finishing work in class time are the areas to grow. At home, agree one participation goal a week and ask them to note how they contributed in a discussion or group task. |
| 9-12 | Strong across the board | [Name] shows mature understanding and applies course concepts well in independent work. They support claims with evidence, meet deadlines and contribute constructively in class. The next step is precision: replacing general words with exact ones. At home, ask them to revise one paragraph with that single aim. |
| 9-12 | Strong, one clear gap | [Name] performs strongly in mathematical reasoning, choosing efficient methods and explaining solutions clearly. Longer written responses are less confident, mainly in how they are organized. At home, ask them to outline the main point, two supporting details and a conclusion before drafting any extended answer. |
| 9-12 | High effort, skill lagging | [Name] consistently puts in the effort, acts on feedback and persists through difficult coursework. Analysis is still developing, especially connecting a piece of evidence to a wider conclusion. At home, have them annotate one source with the claim, the evidence and why it matters, then talk through how the parts connect. |
| 9-12 | Behaviour is the concern | [Name] does thoughtful work when focused and brings a useful perspective to individual tasks. Attendance, preparation and active participation in class are the concern this term. At home, review the next day's schedule together, get materials ready the night before, and set one specific participation goal. |
Every one of the sixteen came back with the same skeleton (strength, gap, home step) because the prompt asked for it, and that is the point of asking: a comment bank with one shape is easy to check and easy to tell apart from the child in front of you. The length is deliberate too. Monsha's generator defaults to a 500-character limit because that is what many report card systems accept; the sixteen above run 262 to 323 characters, so they fit with room for the sentence you add.
How do you edit an AI draft so it reads like you wrote it?
Add the one fact the model could not know, then cut the word it always reaches for. The first is a concrete moment from the term (the fraction wall, the science fair question, the day they helped a new classmate); the second is "demonstrates", which opened six of our 19 first drafts. A parent reads thirty comments a year and notices when every child "demonstrates strong" something.
Then read the batch, not the comment. Georgetown Day's critics were not reacting to any single comment; they were reacting to sameness across a class. Sort your drafts by the growth sentence and rewrite any two that share it word for word. Keep the tone you asked for: the nine run drafts, asked only for "warm and direct", came back with no exclamation marks and no "amazing", so if those creep in, the prompt drifted.
Where the rules live matters when you are doing this for 90 students in a week. On Whizi the comment rules, the tone, the word cap and the "[Name], never a real name" instruction sit in a project's instructions (up to 32,000 characters, any signed-in account) with your rubric pinned as a file, so each new chat starts with them loaded, and the same notes can go to GPT, Claude or Gemini without retyping. Whizi has no gradebook or student information system connection, so unlike Monsha's class list upload or ReportCards.ai's Google Classroom import, nothing about a real student ever enters it unless you type it, and nothing writes back to the report card either. The Projects guide covers the setup; how to use AI for writing covers the generic-prose problem in general, and how to use ChatGPT for homework is the page to hand the students.
- Write the prompt as grade band, subject, three observations, tone and length; ask for three versions, not one
- Keep the student's name, student number, grades and any diagnosis out of the chat; FERPA lists the name first among identifiers
- Check whether your district names an approved tool; Varsity Tutors and Knowt require a name, so a personal login to either sends the name to a vendor your district has not contracted
- Turn off "Improve the model for everyone" in ChatGPT Data controls if you use a personal account at all
- Edit every draft: add one moment from the term, cut "demonstrates", and rewrite any two comments that share a growth sentence
- Tell a parent who asks that you drafted from your own notes with AI and edited every line, and make it true
Frequently asked questions
Can teachers use ChatGPT to write report card comments?
Yes, if the district allows it, and with the student's name and grades left out of the prompt. Kearny Schools in New Jersey gives teachers three LLM platforms for the job as of August 2026, the NEA's sample board policy asks districts to tell parents how AI is used, and the ETFO tells Ontario members to use only approved tools. A personal ChatGPT account trains on your chats unless you turn "Improve the model for everyone" off.
Is there a free AI for report card comments?
Yes. Khanmigo's generator is free for teachers with no limit, funded by Microsoft; MagicSchool is free for individual teachers; ReportCards.ai gives 50 feedback generations free; and the prompt on this page runs on any free chatbot. The cost of the model itself is trivial: our nine-comment run across GPT, Claude and Gemini cost $0.015 in total on 27 September 2026.
Can AI personalize report card comments for each student?
Only as far as your notes go. Given three observations the models on this page wrote three distinct comments each, all true to the notes; given a grade and a name they would write the same comment for every child with that grade. The personal part is the observation you type and the one moment from the term you add while editing.
How long should a report card comment be?
Between 40 and 70 words for most systems. Monsha's generator defaults to 500 characters because that is a common report card field limit, and the sixteen edited comments in our table run 262 to 323 characters, while the three unedited run comments shown run 56 to 60 words. Ask the model for a word range in the prompt; all three models we tested held a 50 to 70 word cap on every comment.
