Math Self-Study Module

I. TL;DR — What to Build This Week

Build a P6 math self-study prototype this week. The module should do three things generic AI tools cannot: (1) pick questions aligned to HK curriculum by skill prerequisite graph, (2) track per-student mastery across sessions, (3) give the teacher a dashboard of who's struggling and where. Skip gamification, skip streak mechanics, skip leaderboards. Focus on the post-wrong experience — that's where the competitive whitespace is widest and teacher demand is strongest.

The prototype should be demoable at P6 level because Essai already has 100 verified P6 questions in DB, teachers explicitly said P6 students struggle most with self-study,¹ and KooBits (the dominant HK primary math platform) stops at P6 with no AI tutoring layer — making it the exact transition point where Essai can differentiate.

II. Feature Recommendations — P0 / P1 / P2 / P3

P0 — Must ship in prototype (this week)

P1 — Ship in v2 (2-3 weeks)

P2 — Ship when school signs (1-2 sprints)

P3 — Defer or cut

III. Student Pain Points — Why Self-Study Fails

The research reveals a consistent pattern across platforms and geographies. Math self-study fails for three reasons, always in this order:

Pain 1: "I got it wrong and I don't know why"

The #1 failure mode. Students answer incorrectly, see the right answer, and still don't understand what went wrong. Drill apps that show only correct/incorrect create a learned helplessness loop.⁹ i-Ready is the canonical failure: parents report children going from loving math to hating it because the app provides no meaningful post-wrong support.⁵

Pain 2: "I keep doing the same easy stuff and it's boring"

Non-adaptive platforms serve the same difficulty level regardless of student ability. Advanced students are bored; struggling students are overwhelmed. The "goldilocks zone" — questions just beyond current ability — is where learning happens.² Math Academy's diagnostic places students at their exact "knowledge frontier" in 30-45 minutes, avoiding both problems.¹⁰

Pain 3: "I practiced a lot but forgot everything"

Without spaced repetition, math practice has a half-life. Self-learners report a frustrating cycle of forgetting material after breaks, having to re-learn foundational concepts.⁹ Math Academy's FIRe algorithm addresses this by giving fractional credit for implicit prerequisite practice — when you solve a hard problem, you're also reviewing the easier skills it builds on.⁷

IV. Competitive Landscape — 12 Players Mapped

Global Players

HK-Specific Players

V. Why Not Just Ask ChatGPT? — The 3-Layer Moat

Every school considering Essai will ask: "Can't students just use ChatGPT?" The answer is yes, they can — and it won't work for math practice. Here's why, with evidence:

Layer 1: Curriculum alignment

ChatGPT doesn't know what P6 students in HK should be learning this term. It can't distinguish a TSA-level question from a DSE-level question. Essai's question bank is tagged by grade, topic, difficulty, and curriculum standard — every question maps to the HK syllabus. This is table stakes for school adoption, and it's something generic AI simply cannot do.

Layer 2: Progress tracking across sessions

ChatGPT has no memory between sessions (unless manually configured). It can't tell you: "This student has attempted 15 trigonometry questions and consistently fails on sin/cos ratio applications." Essai's atomic skills tracking persists across sessions, building a mastery profile per student. This is what powers adaptive question selection.

Layer 3: Teacher visibility and control

This is the purchase decision layer. Teachers need to see which students are struggling and on which skills. They need to assign specific practice. They need to disable AI help for exams. None of this exists in ChatGPT, Claude, NotebookLM, or Gemini. It's the institutional layer that justifies school-level procurement.

Where GenAI tools ARE useful

Don't fight GenAI — integrate it. The AI chat (Ask AI) that Leslie is building should use an LLM for Socratic guidance when a student is stuck on a specific question. The key design constraint: feed the LLM the question context, the student's wrong answer, and the relevant atomic skill, so it can guide without hallucinating. This is exactly what Khanmigo does — and Essai can do it better because Essai owns the question structure.

VI. Grade-Level Recommendation — Start with P6

VII. What the Best Get Right — Design Patterns to Steal

From Math Academy: the prerequisite graph

Math Academy's 3,000-topic knowledge graph is their moat. A third-grader completed six years of math in one year because the system never asks a question the student isn't ready for — it traces prerequisites all the way back.¹² Essai's competency maps (workdir/math/knowledge/competency-map/) have prerequisite links. Use them.

From IXL: the SmartScore simplicity

IXL's SmartScore (0-100) is brilliantly simple. Wrong answers make it harder to advance; correct answers on harder questions earn more. Students always know where they stand. The "Challenge Zone" at 90+ gives advanced students something to reach for.³

From Khan Academy: the Socratic constraint

Khanmigo never gives the answer. It asks: "What do you think the first step would be?" This is a design constraint, not a feature — and it's critical. The Ask AI chat must be Socratic, not solution-providing. Feed it the question's toAce and errorTraps fields so it can guide precisely.

From Mathspace: effort over outcome

Mathspace rewards question streaks and effort, not just accuracy. Students earn more points for sustained focus. Daily and weekly challenges create low-barrier return triggers.¹⁷ This is the minimal-viable engagement loop — simpler than Duolingo's full gamification stack, but effective.

From the Oral Module: session structure

Essai's Chinese/English oral module has school adoption. The pattern: student enters → picks level → immediate practice → immediate feedback. No complex setup. The math self-study module should follow this exact entry flow, adapted for adaptive question selection instead of fixed prompts.

VIII. HK Market Context — Timing and GTM

EDB AI for Education Funding Programme

Launched December 2025. HK$500 million total, HK$500,000 per school.²⁶ Application deadline: February 28, 2026. Schools must implement AI-assisted teaching in at least 3 subjects and conduct public demo lessons. This is the single biggest GTM lever for Essai right now:

• Schools have budget earmarked specifically for AI education tools
• They need to show AI usage in 3+ subjects — math self-study with AI tutor counts
• Funds can cover subscriptions, devices, and AI-powered services through August 2028
• SmartQuest is already EDB-approved. Essai should pursue the same approval.

Competitive position in HK

• Primary (P1-P6): KooBits dominates with 200K users and elite school adoption. But KooBits is Singapore-method drill — no AI tutor, no Socratic guidance, no post-wrong scaffolding.
• Secondary (DSE): SmartQuest has 80+ schools on free trial with AI marking + paper generation. Essai's self-study mode is complementary (practice vs. assessment).
• Gap: Nobody owns P6 AI-guided self-study in HK. This is the whitespace.

Geniebook's lesson

Geniebook achieved profitability in 2024-25 after pivoting to hybrid (online + physical centers).²² This suggests pure-platform may not be enough for scale in HK — but for Essai's current phase (trial schools, prove the product), pure platform is correct. Hybrid comes later if at all.

IX. Critical Assessment of P0 Choices

P0: Adaptive question selection by prerequisite graph

Source test: PASS — Math Academy's knowledge graph approach is backed by research papers and demonstrated 4x learning speed improvement.²

Feasibility test: PASS — Essai's competency maps already have prerequisite links. The algorithm is deterministic (JS), not AI-dependent. No accuracy gate.

Adoption test: CONDITIONAL — Adaptive selection only works if students actually use the module repeatedly. Essai's B2B2C model helps: teachers assign practice, students have to do it.

Counter-example: i-Ready is adaptive but universally hated.⁵ The difference: i-Ready makes students listen passively. Essai's approach is active problem-solving. The key is what happens after the question, not the question selection itself.

Reformed position: HOLD — Ship it. The existing competency maps make this low-effort, high-impact.

P0: Post-wrong hint escalation

Source test: PASS — Every successful math platform (IXL, Khan, Mathspace) has some form of post-wrong support. Absence of this is the i-Ready failure mode.

Feasibility test: PASS — Leslie already built hint infrastructure. Eric's questions have errorTraps and toAce fields. Progressive reveal is straightforward.

Counter-example: Photomath gives full solutions immediately — and is criticized for enabling copying, not learning.¹⁴ The progressive reveal pattern (hint → step → solution) prevents this.

Reformed position: HOLD

P0: Teacher visibility

Source test: PASS — Mar 15 sync: teacher explicitly asked to monitor AI tutoring sessions.¹

Adoption test: PASS — This is the B2B purchase decision feature. Schools buy tools that give teachers visibility. Without this, Essai is just another practice app.

Reformed position: HOLD — Even a minimal version (list of students + skills attempted + accuracy %) is valuable for the prototype demo.

X. Live Signals — What Practitioners Are Saying

XI. Open Questions

1. Cantonese AI chat: No major AI has good Cantonese math tutoring. Is English acceptable for the AI tutor layer, or does this block adoption in HK primary schools?
2. Leslie's Ask AI timeline: The AI chat component is Leslie's scope. When does it ship? The prototype can work without it (hint-only), but the AI tutor is the differentiator for demos.
3. EDB approval process: SmartQuest is already EDB-approved. What's the application process? Is there a vendor list? Who at Essai handles institutional relationships?
4. Spaced repetition complexity: Math Academy's FIRe algorithm is sophisticated. For v1, is simple "review skills you got wrong in the last 3 sessions" sufficient?
5. Session length expectation: How long should a self-study session be? Oral module sessions are ~5-10 minutes. Math may need 15-20 minutes to be useful. What do teachers expect?