2026-05-14 · ActiKidz Team

AI class reports: turning lesson notes into parent-ready updates in 30 seconds

How much time tutors actually spend writing reports

Rough math: a tutor with 30 students, writing one report per student per week, 5–8 minutes per report (recalling lesson details, then translating into something a parent can read) — that's 2.5–4 hours a week of pure writing.

Four hours that could go into prepping next week's lessons, onboarding two more students, or just having dinner with your own kids.

The obvious first idea is "build a template and fill in the blanks." We tried that. About half an hour later we realized it doesn't work — "Today XX did well, please keep practising" is something parents can spot from a mile off. A templated report is worse than no report, because it tells the parent "I don't care about your kid."

So the problem isn't saving time. It's saving time without losing personalization. That's not as simple as it sounds.

Where AI actually fits

It took us a few iterations to figure out — AI doesn't belong at the front (writing for the teacher) or at the back (sending for the teacher). It belongs in the middle: rewriting the teacher's internal notes into something parent-facing.

The workflow ends up looking like this:

1. Teacher spends 60 seconds jotting down 5–8 notes — abbreviations, shorthand, anything
   • e.g., "right H not lifting, rhythm ok, Hanon 21 line 2 stuck 3x, needs slow practice"
2. AI takes 30 seconds to generate a paragraph —
   • "In today's lesson Emma focused on Hanon 21, a foundational technique exercise. Her overall sense of rhythm was solid, but her 4th finger on the right hand (the one most beginners struggle to move independently) needs more strength. For homework, please practise line 2 slowly and focus on a clean, deliberate strike for each finger."
3. Teacher spends 30 seconds editing: add a personal encouragement, fix a term, append next lesson's plan
4. Send

Total time drops from 5–8 minutes to about 2 minutes. 2–3 hours saved per week.

The key thing is that the teacher's judgment is still in the loop — what to note up front, what to fix at the end. AI only handles the rewriting part, which is the most mechanical part of the job. It's basically a translator turning teacher jargon into parent-readable language.

Things we thought hard about while writing the prompt

Throwing teacher notes at OpenAI sounds simple, but the landing is full of small traps. Below are a few decisions we spent real time on — not claiming any of them is optimal, but each one came from actually getting it wrong first.

We didn't build "K–3 / 4–8 / 9–12" three templates

Originally we wanted separate prompts per age group — one for K–3 piano, one for 9–12 academic. Before we even started building, we hit a problem:

Age group is only one dimension. There's also discipline (piano vs. math), parent style (data-driven vs. feeling-driven), teacher style (structured vs. narrative). Cartesian product gets you a dozen-plus templates, heavy ops overhead, and a redeploy every time you change a sentence.

We ended up with one FreeMarker template + field variables. The prompt is assembled dynamically:

ageGroup: "8-12"           
toneStyle: warm | formal | encouraging | professional
teacherPreferences: { ... }
observations: ["right H not lifting", "rhythm ok", ...]

Adding a new tone style doesn't need a deploy — you just edit a field config in the admin panel. That flexibility has saved us several times since.

Default to gpt-4o-mini, not gpt-4o

OpenAI's current public pricing (2025):

30 students × $0.00036 = $0.011 per week, $0.55 per year.

We blind-tested 50 reports comparing gpt-4o-mini and gpt-4o output — tutors literally couldn't tell which was which. But the cost difference is 16×.

So gpt-4o-mini is the default. We kept a preference toggle for "use a stronger model" (included with the PRO plan), but the default is enough for most tutors.

If you're building a similar "AI rewrite" application, start with the cheapest model. Most tasks don't actually need much model reasoning — try mini first, upgrade only if it's genuinely not enough.

max_tokens=1000, temperature=0.70

• max_tokens=1000: A report in English runs 200–400 words; 1000 is more than enough. But the hard cutoff is a safety net — if the AI decides to ramble, one generate call won't burn through your budget on a 10,000-word essay.
• temperature=0.70: 0.3 is too mechanical (30 reports start reading like the same person wrote them), 0.9 is too random (style drifts every call). 0.70 is the value we landed on after a few rounds of tuning.

Both parameters live in the ai_prompt_templates table, so ops can tune them per tutor or per template without a redeploy.

observations capped at 10, truncate if more

Sometimes tutors get carried away — "she came in saying she was tired, so I let her rest for 2 minutes, then she said her mom yelled at her this morning..." But the longer the prompt, the more the AI's attention diffuses, and quality actually drops; token cost is also linear.

We hard-cap at 10 notes (truncate by recency). Early on a tutor pushed back: "I wrote 20 and you cut them." Our response was: if you have 20, you probably need two reports covering different themes, not one.

24-hour cache

Tutors sometimes generate once for preview, edit, then generate again. Observations didn't change. We cache these repeats for 24 hours — same hash, same result, no API call.

Pre-cache OpenAI bill was $30/month. Post-cache: 2/month. 60% drop.

This kind of cost optimization is boring, but it works.

Timeouts at connect=5s / read=15s

OpenAI hiccups occasionally (especially peak hours). With the default 120s timeout, tutors assume the system is broken and start spamming the retry button — which triggers more API calls and makes everything worse.

We use 5s + 15s short timeouts. If we don't get a response in 15 seconds, we fall back to "AI temporarily unavailable, please try again." The tutor's workflow doesn't lock up, and OpenAI's side doesn't get stacked retry pile-ups.

Things we got wrong

Wrong 1: tutor abbreviations went straight to parents

Week one of launch. A parent received: "Today right H not lifting, Hanon 21 line 2 stuck 3x."

The parent had no idea what that meant.

Fix: amend the system prompt to add — "Your output is for parents who don't understand musical terminology. Any abbreviation, internal shorthand, or technical term must first be explained in plain language, with the original term in parentheses."

After that, "H not lifting" became "the 4th finger on the right hand (piano fingering notation H) needs more strength." Parents could actually read it.

This reminded us of an old lesson: in a system prompt for an LLM, "what it doesn't know" matters more than "what it should do." AI won't proactively assume things about its audience — you have to tell it.

Wrong 2: default toneStyle was too stiff, like a court summons

Early on the default was professional. Output looked like this:

"Per evaluation, the student achieved 75% of the established lesson objectives for this session."

Parents universally complained it "reads like a lawyer's letter." One parent literally asked "are you trying to dismiss my kid from the program?"

We switched the default to warm:

"Emma completed most of the required practice goals in this session, with clear overall progress."

Same information, completely different reception.

We learned something useful here: the default is the setting most users will never change. So the default should map to "the version most parents will respond to best in most situations" — not "the most formal version in theory."

This rule applies pretty broadly. Every SaaS default should be picked this way.

Wrong 3: student names got jumbled

Student records have nameZh: 张小明, nameEn: Tom Zhang, displayName: Tom. The AI received a note like "张小明 today..." but the report needed English output. So it transliterated on its own: "Zhang Xiao Ming."

The parent didn't know who the teacher was talking about.

Fix: pass displayName explicitly into the prompt, and add to the system prompt — "Refer to the student only using the provided displayName. Do not transliterate, do not mix Chinese and English."

Small details like "use exactly the field value you were given" make a big difference in feel.

When you should not use AI

Not every report should be AI-written. The cases where we recommend tutors write by hand:

1. Major regressions or behavioral issues — AI tends to produce "challenges and opportunities coexist" platitudes, which parents read as evasion and which deepen anxiety.
2. Parents who've previously complained — any sensitive family conversation lands better when it's clear the tutor wrote it personally.
3. Complex family backgrounds (divorce, immigration anxiety, other particulars) — generic fluency comes across as cold here.
4. Major milestones (first time passing Grade 8, making Math Olympiad finals) — worth 10 minutes of the tutor's own writing, not 30 seconds of AI.

That's why the ActiKidz AI class report is opt-in — for every report, the tutor can choose "AI draft" or "fully manual." Never forced. AI is a tool, not an assembly line.

Numbers

• 30 students/week, tutors save 2–3 hours/week on report writing
• Monthly OpenAI cost: 2–15 (after caching)
• Per-report generation: 15–20 seconds (includes API call + DB lookup + FreeMarker render)
• Tutor acceptance rate (generated reports that actually go to parents): about 85% — the other 15% are regenerated or rewritten by hand

The ActiKidz AI class report feature is live, included in all paid plans. Tutors can adjust default tone, customize preferences, and view monthly token usage from the "AI preferences" panel in the dashboard.

tl;dr

In the tutoring use case, AI should do the mechanical rewriting, not the writing itself. The tutor's eyes, judgment, and care stay where they belong — observing the student up front, polishing the wording at the end. The 30 seconds in the middle, where words get reorganized, is what you hand to AI.

This isn't a romantic "AI revolutionizes education" story. It's a boring, practical 2–3 hours saved per week. Good enough.

Other domains (medical, legal, sales debriefs) might split human and AI differently. But for tutoring, our gut feel is this: humans at the front and back, AI in the middle.

Want to try it: pricing / contact sales.