§ Findings
What shows up across the interviews
1
The AI tutor solves a real classroom problem, not a convenience one.
This is the strongest product-market-fit signal in the three sessions. Both Adif and Afira described their real-world practice partners as the bottleneck — not their own motivation, not access to materials, not their teacher.
My speaking partner — she doesn't like to talk to me about English stuff because she's also struggling with it. So Roly — it's like speaking to me. I like that.Afira, Form 5
The contrast with Dania makes it sharper: she doesn't even have a regular practice partner — she "sometimes" does speaking with her teacher in class. Same shape of need, more acute.
Implication · This is the marketing line. The AI tutor isn't replacing a willing human partner. It's filling the gap left by an unwilling, unprepared, or absent one.
2
There's real format uncertainty about Paper 3 — even among students sitting it.
The three accounts of what Speaking Paper 3 actually looks like don't agree. This is itself a finding.
| Element | Adif | Dania | Afira |
|---|
| Number of sub-parts | 4 | Affirmed when asked | 3 (Part 4 = Listening Paper) |
| Part 1 length | "Q3-5 are random" | — | Name + 1 question only |
| Part 4 of Speaking | Decision-confirmation dialogue | — | Didn't mention |
Two readings: either the format genuinely varies by examiner and school, or one or more students are mis-recalling. Either way, the app shouldn't lock to one exam format.
Action · Get the Lembaga 1119/3 official spec from Cikgu Natasha directly before any more content writing. The teacher knows definitively — students are unreliable narrators of exam mechanics.
3
Students want more scaffolding, not less.
A surprising and consistent ask across both engaged sessions.
On part three, you guys can give a template — a template answer to student to use so that the student doesn't clueless what to answer it.Adif
Roly chose two and I only had to choose one. So it's very easy for me to elaborate. It lessens my decision making.Afira, on Part 3
Adif also described how Cikgu Natasha actually preps her class: "We just remember the template, not the topic, because the topic is actually different." Students don't memorise day-specific answers — they memorise structural skeletons they apply on the day.
Implication for the 30-day backlog · Every Part 3 day should include an explicit template — a 4–5-line skeleton (open → reason → invite partner → respond → decide). The model arguments fill the skeleton; the skeleton is what transfers.
4
Part 1 is more flexible than our spec assumes.
Adif described Part 1 as "Q3–5 are random" — openers fixed (name, where from) plus a few rotating fillers. Afira said the real exam is just name + 1 question. Both said the longer app version (currently 6 questions) is helpful for revision but doesn't mirror the real test.
Resolution · Default mode asks 3–5 questions with rotation (matches Adif's reality). Add an exam-mode toggle that strips down to 1 for the days right before the trial / real SPM.
5
Adif and Afira are the same persona. Dania is a different one.
The most consequential strategic finding from N=3. Two students look almost identical in needs and behaviour; the third sits in a different segment with different anxieties.
Persona A · High-engagement
Adif & Afira
- B1+/B2 fluency, can self-narrate in full sentences
- Anxious about Part 3 speaking specifically
- Independently engaged with the prototype
- Use the speaking module purposefully
- Found via teacher recommendation
Persona B · Lower-fluency
Dania
- B1 or below, one-word answers in interview
- Anxious about essay (Paper 2) and vocabulary
- Couldn't access the speaking prototype (mic + fluency floor)
- Already uses an Instagram English account daily
- Discovers tools via her feed, not teacher
Strategic implication · The Paper 3 prototype is well-targeted for Persona A. For Persona B, the product probably needs a different surface — lower-friction, vocab-first, essay-prompt-led, discoverable from Instagram. Worth a 15-minute conversation with Isaac about whether to keep both personas in scope.
6
The character-led UI is doing real emotional work.
All three students made some version of this comment, unprompted:
It's so cute … reminds me of Shopee.Afira
One of the best design on our educational app. The layout is simple, clean and easy to use. And this app does not contain any pop-up advertisement.Adif
The photo is cute, maybe.Dania, in a session where she barely commented on anything
Three unprompted positive responses across three different proficiency levels.
Implication · Keep the warm-orange palette and character-led UI. They're paying off. And keep the no-ads stance Adif flagged — for him it was a trust signal, not just a hygiene one.
§ Backlog changes
Three updates to fold in before writing more days
Add a Block 3.5 — Decision lock.
Adif's Part 4 detail: 30–60 sec of model dialogue per day where the AI tutor and student formally agree on one option and one reason. Currently missing from the Day Structure spec entirely.
Add a "Today's template" mini-section above each Part 3 day.
4–5 lines of structural skeleton, not topic-specific. Mirrors how Cikgu Natasha actually preps the class. Students transfer the template, not the answer.
Soften Part 1 from 6 fixed questions to 3–5 with rotation, by default.
Build the exam-mode toggle that strips down to 1 for the days right before the test.