Progress reports and the feedback ladder

The inputs the model actually needs

The reason most AI-drafted report comments come out generic is that the teacher gives the model nothing specific to work with. “Write a report card comment for a Class 6 maths student” will produce sixty words of pleasant nothing every time. The model has no facts, so it falls back on the average of every school-comment ever written on the internet — which is, by design, no school in particular.

The fix is the same fact-first discipline you used for parent updates, scaled down to fit one student. For each child, before you ask for a comment, give the model a compact factual block:

• Marks. The actual numbers — term marks, exam marks, or marks in the relevant component. Not “good” or “average.” Numbers.
• Attendance. A figure or a short pattern. Eighty-eight percent, two unexplained absences in Mangsir.
• One or two lines of teacher evidence. This is the load-bearing input. A specific observation only you have, that no register can produce. “Struggles with word problems but is the strongest in mental arithmetic in the class — keeps getting full marks on the daily ten-question warm-up.”
• The tone you want. Warm-but-honest works for most parents. Some parents — and you will know which — need a slightly firmer register; some need a softer one because the child is fragile right now.
• The audience. A line about whether the family will read this in Nepali, English, or both, and whether the guardian is the parent, a grandparent, or an older sibling.

That is the entire prompt: five or six lines per student. The teacher evidence — the line only you can write — is what stops the model from producing a comment that could be about any child in the country. Without it, the model is guessing. With it, the model is wrapping your knowledge of this child in language the family will read.

The structured comment: one strength, one next step

A good report-card comment, in any country, has exactly two parts: one specific strength and one concrete next step. Anything more is decoration; anything less is unhelpful.

The specific strength is not “works hard” or “is enthusiastic.” It is “in the Magh project on fractions, Sunita was one of the few students who could explain why two-quarters and one-half were the same — she is starting to see the reasoning behind the rules, not just the rules.” The parent reads that, the child reads that, and something happens in both of them. The parent realises this teacher actually saw their child. The child realises she has been seen doing something she did not know was visible.

The concrete next step is not “should pay more attention” or “needs to study harder.” It is “the next step is word problems — Sunita can do the calculation, but she still freezes when the question is given as a story. Practising one or two word problems aloud at home each evening, with the family asking ‘so what is the question really asking?’, would close this gap quickly.” The parent now knows what to actually do. The child now knows what the teacher means by “improvement.” The advice is actionable in this exact household.

Two parts. Three or four sentences. No generic praise. No empty exhortation. The model produces this shape effortlessly once the facts and the structure are specified, and refuses to produce it when they are not. The discipline is in the prompt, not the model.

The feedback ladder

Here is the test a comment must pass to be worth the ink. If the parent reads it out at dinner and asks the child “what does your teacher mean by that?” — can the child point to a specific thing?

This is what we will call the feedback ladder. Comments that pass it sit on a real rung — a specific lesson, a specific test, a specific moment in the classroom that the child can recall. Comments that fail it float in mid-air — they refer to nothing the child can recognise, so the conversation at home dies, and nothing changes by next term.

The ladder also explains why AI-drafted comments often feel better but do less. A well-written generic comment is more pleasant to read than a clumsy specific one, but a clumsy specific one drives behaviour and a polished generic one does not. The goal is not pleasant; the goal is acted-on. Specificity is the only path. AI’s role is to take your specific input and make it pleasant as well as specific — both, not either.

A worked example: thirty Class 6 maths comments

The realistic workflow looks like this. You sit down on Friday evening with your Class 6 maths register, the term test paper, and one cup of tea. The register has thirty names, the marks for each, and the attendance figure. Beside the register you keep an open notebook in which, over the term, you have jotted one or two short observations per student. “Anil — mental arithmetic strong, copies in word problems.” “Pratima — slow in class but produces beautiful diagrams.” “Rabin — can do everything but rushes and loses marks on copying.” These notes are the gold; without them the next forty minutes do not work.

You open Claude or ChatGPT. You paste a single prompt template once: “For each student below, write a two-to-three sentence report comment in Nepali and English. Use only the facts given. Follow this structure: one specific strength, one concrete next step. Tone: warm but honest. End the comment with the next step phrased as a small thing the family can do at home.” Below the template, you paste your thirty fact blocks, one per student, three lines each: marks, attendance, your one teacher-evidence line.

The model returns thirty drafts. Each one names a specific thing, suggests a specific home action, and is in both languages. You spend about forty seconds per student reviewing — not editing every word, but checking the three things that matter: does this sound like me, does it match what I actually saw, and is the home action realistic for this family? You tweak perhaps eight of the thirty. You reject perhaps two and rewrite them by hand because the model picked up the wrong angle. The whole batch is done in fifty minutes. Last term, the same batch took four hours.

The difference is not that the model is smarter than you. The difference is that the model is faster at language and you are faster at judgement, and the workflow gives each what it is good at.

The teacher review: thirty seconds per child

The final step is not optional. Every AI-drafted comment gets a teacher pass before it goes into the report card. The pass is short — thirty seconds — and it asks three questions.

Does this sound like me? If the comment uses words you would never use — “showcase,” “exemplary,” “leverage” — strike them and replace. The parent who has read your handwritten comments for two terms will notice the moment you start sounding like a different person. They may not say anything, but a small bit of trust leaks out, and trust is the whole basis of the parent-teacher relationship.

Does this match what I actually saw? The model can subtly overstate. A child who is “improving slowly” can become “showing strong growth” if the prompt nudges that way. Bring it back. The comment must be defensible: if the parent walks in next week and asks you to show evidence, you must be able to do it without flinching.

Is the next-step action realistic for this family? “Practising word problems aloud at home each evening” assumes a literate adult at home with twenty spare minutes after dinner. That is not every family. For a child whose grandmother is the at-home guardian and whose parents are abroad, you might shift the action to “five minutes of mental arithmetic on the way to school with friends — Anil already does this naturally and benefits from it.” The model will adjust the moment you tell it the family situation.

Thirty seconds per child, thirty children, fifteen minutes total. After fifty minutes of drafting and fifteen minutes of review, you have a stack of report comments that will move something in each family, and your Saturday is your own again.

Check your understanding

Quick check

—

You sit down to write thirty Class 6 maths report comments. Which workflow produces comments that actually mean something to parents and children?

1. AAsk the model — write thirty warm report card comments for a Class 6 maths class — and copy whatever it produces straight into the report cards.
2. BFor each student, give the model a fact block (marks, attendance, one specific teacher observation), ask for one specific strength plus one concrete next step in both Nepali and English, and then spend about thirty seconds per comment reviewing for voice, accuracy, and whether the home action is realistic for that family.
3. CPaste your full mark sheet and attendance register, including every student name and roll number, into a public chatbot and ask for a batch of comments.
4. DWrite all thirty by hand the way you always have, and never use AI for anything related to reports.

Check answer→

What comes next

Report cards are the visible work. Behind them sits the invisible administrative work that quietly eats whole evenings — meeting minutes, sports day notices, scholarship forms, school-day circulars, exam schedules. The next section is about templating those once, with AI, and then reusing the templates mechanically across the term — and the data-discipline rule that decides what may and may not go into the chatbot to do it.