Khurram Badar / Archive / Papers / newworld.education — Architecture Brief

newworld.education — Architecture Brief

other · 2026-04-19 · 4639 words · Khurram Badar

newworld. education — Architecture Brief.

ai · education · web

newworld.education — Architecture Brief

**For:** Brain Games + Platform Education Layer + Starky Upgrade
**Audience:** Khurram (review + approval) → Claude Code (execution)
**Version:** v1 — draft for review, not yet final
**Status:** Architecture only. No code is to be written until this is approved.

---

0. Read this first

This brief architects three systems that work as one. Do not build them in isolation.

They share one student, one progress model, one parent view, one Starky. Anything that breaks that unity is wrong by default.

This is an architecture document, not an implementation plan. It tells Claude Code what to build and what not to build — not how to write every line. Claude Code fills in implementation details. Khurram reviews every major decision before it ships.

---

1. The strategic thesis

newworld today is a test-prep product competing in a crowded market on question volume. The upgrade repositions it as an **outcome-driven learning platform** where three things compound together:

1. The **core curriculum layer** teaches what the child needs to know for Cambridge / Edexcel / local syllabi.
2. The **Brain Games layer** strengthens the underlying cognitive skills that make learning faster — memory, attention, pattern recognition, spatial reasoning, decision-making.
3. **Starky** connects the two in real time — when a child struggles on a math problem, Starky notices the underlying cognitive gap and prescribes a Brain Games drill; when a child completes a Brain Games streak, Starky shows the parent where it translated into exam performance.

No competitor in the GCC / South Asia education market currently offers this triangle. E2Language doesn't. Cambridge's own tools don't. Khan Academy doesn't. The moat is the integration, not any individual layer.

**Pricing implication:** the current $9.99/month is no longer a test-prep subscription. It is a child-development subscription. Parents pay for measurable change in their child — grades, focus, confidence — not for more practice questions. This changes what "retention" means: a parent who can see their child getting measurably smarter will not cancel.

**Market fit:** GCC HNW families, Pakistani and South Asian diaspora, aspirational middle-class across emerging markets. Parents who value provable development over vague engagement metrics. Not casual edtech consumers.

---

2. Audience and voice (non-negotiable)

Primary user: the student (age 5–18)

Voice must adapt to age:
- **KG to Year 2 (5–7):** playful, lots of Starky encouragement, images over text, celebrate every step, no numeric scoring visible
- **Year 3 to Year 6 (7–11):** game-like, achievements visible, Starky is a friendly coach, curiosity-forward
- **Year 7 to Year 9 (11–14):** skill-building framing, Starky is a study partner, show progress clearly, competitive but healthy
- **Year 10 to A Level (14–18):** serious mastery framing, Starky is a tutor, exam-relevant, no babying

Secondary user: the parent

Parents want three things and nothing else: *Is my child learning? Is my child getting better? What should I do about it?* The parent dashboard answers exactly those questions in plain language. Not dashboards of metrics. Sentences.

Voice DNA

---

3. Brain Games — the cognitive training layer

3.1 Philosophy

The National Geographic *Brain Games* book by Jennifer Swanson uses one brilliant pedagogical pattern that we adopt (but do not copy the content of):

> **Challenge → Experience → Explanation → Application**

The child does something, feels what happens, learns why their brain did that, then applies it to real schoolwork. This four-beat loop is the pedagogical spine of every Brain Games module. It is also what every other cognitive training app on the market is missing — they do the challenge and skip the other three beats.

3.2 The five modules

Each module is a distinct cognitive domain with its own games, its own scoring, its own progression. They are not five skins of the same puzzle engine. They exercise genuinely different parts of the brain.

**Module 1 — Memory Lab**
- Exercises: short-term recall, long-term encoding, spaced repetition, associative memory, working memory
- Real curriculum bridges: periodic table, verb conjugations, historical dates, Quran / scripture memorisation (opt-in), formulas, vocabulary
- Starky framing: "You just held seven things in your head at once. Here's how that helped you on yesterday's chemistry quiz."

**Module 2 — Pattern Recognition**
- Exercises: number sequences, visual pattern completion, analogy chains, rule induction, anomaly detection
- Real curriculum bridges: algebra, geometry, science hypothesis formation, grammar patterns, music theory
- Starky framing: "Spotting that pattern is the same skill you need for quadratic equations. Want to try one?"

**Module 3 — Focus Training**
- Exercises: Stroop-style conflict resolution, sustained attention, dual-task management, distraction resistance, reading comprehension under time pressure
- Real curriculum bridges: exam stamina, long-reading passages, multi-step word problems
- Starky framing: "Cambridge A Level exams require two hours of focus. This drill builds exactly that muscle."

**Module 4 — Spatial Reasoning**
- Exercises: mental rotation, 3D visualisation, map-reading, graph interpretation, mirror and symmetry
- Real curriculum bridges: geometry, physics (forces, optics), chemistry (molecular structure), biology (anatomy), geography
- Starky framing: "Rotating that shape in your head is what chemists do with molecules every day."

**Module 5 — Decision & Logic**
- Exercises: risk/reward scenarios, logical fallacy spotting, argument mapping, Bayesian reasoning (age-appropriate), ethical dilemmas
- Real curriculum bridges: essay writing, critical thinking, science method, debate, economics
- Starky framing: "You just picked the better long-term choice. That's the same thinking that wins essay marks in English."

3.3 The three-layer rule

Every Brain Games session must have all three layers or it is not shipped:

1. **The game itself** — genuinely fun, self-contained, completable in 2–5 minutes.
2. **Starky's explainer** — "here's what just happened in your brain" — age-appropriate, honest, curious, 30 seconds of reading.
3. **The curriculum bridge** — "here's how this connects to what you're studying" — shown automatically when the student has recent curriculum activity; hidden for young learners who don't yet have curriculum history.

A Brain Games session without the curriculum bridge is just a puzzle app. The bridge is the entire moat. Do not ship without it.

3.4 Content generation

All Brain Games content is generated by Claude Haiku at session time. No hand-authored puzzle bank. No licensed content. No scanned images from any book. The cost advantage (roughly $0.001–0.003 per session) is what keeps unit economics at 85%+ gross margin.

Each module has a master prompt (written once, versioned, committed to the repo) that instructs Haiku how to generate puzzles in that domain at the appropriate age and difficulty. Master prompts include:

**Critical:** Puzzles that a student answered in the last 30 days must not repeat. Haiku must be given a "recent puzzle hash list" and instructed to generate something new. Implementation: store a hash of each puzzle Starky generates per student; pass the last 200 hashes back with every generation request.

3.5 Scoring and progression (non-obvious design)

Brain Games does not use XP / levels in the traditional game sense. Those feel hollow to parents. Instead:

3.6 What Brain Games is NOT

---

4. Platform Education Layer — the structured learning upgrade

4.1 What is broken today

newworld currently has 89,000+ questions across 30+ subjects. That is an impressive question bank but a weak learning product. Questions without a journey are drills. Students and parents cannot answer "what should I do next?" or "am I on track?"

4.2 What replaces it

A three-tier structure that wraps the existing question bank without discarding it.

**Tier 1 — Subjects** (e.g., Cambridge IGCSE Mathematics)
**Tier 2 — Units** (e.g., Algebra, Geometry, Statistics) — pre-defined per exam board
**Tier 3 — Lessons** (e.g., Solving Quadratic Equations) — granular, 10–20 minute learning units

Each Lesson contains:
- A short **Explainer** — Starky teaches the concept in 2–5 minutes, with examples (text + optional diagram; audio narration optional for younger learners)
- A **Worked Example** — Starky solves one problem end-to-end, showing the thinking
- **Practice Problems** — 5 to 15 questions drawn from the existing 89k bank, selected by difficulty and tag
- A **Mastery Check** — a short adaptive quiz that determines if the student has mastered the lesson (gate to next lesson)
- A **Review** — Starky summarises what was learned, flags any weak spots, suggests a Brain Games drill if cognitive gap detected

4.3 Learning Paths

A Learning Path is an ordered sequence of Lessons designed to move a student from their current level to a specific goal. Example: "IGCSE Mathematics — Full Syllabus, exam in June 2027." The platform auto-generates a recommended path based on:

Parents and teachers can override paths. Students cannot, to prevent cherry-picking easy topics.

4.4 Diagnostic Assessment

This is the most important onboarding moment and the hardest to get right. It is a single 20-minute assessment that:

**Critical:** The diagnostic must not feel like an exam. Starky frames it as "let me figure out where you are so I can help you properly." Failure must not feel like failure. No red X marks. No score shown until it is contextualised with a plan.

4.5 Curriculum coverage (scope fence)

Phase 1 launch supports:
- Cambridge: KG, Primary, Checkpoint, IGCSE, AS/A Level
- Edexcel: Primary, IGCSE, IAL
- Pakistani Federal Board (FBISE) — because Khurram's audience is heavily Pakistani diaspora
- UAE MoE curriculum — because Khurram is in Dubai and schools matter

Phase 1 does NOT support: CBSE / ICSE (India), O/A Level Singapore, IB, US Common Core, Saudi national curriculum, African national curricula. These are Phase 2 or later. Do not build scaffolding for them now — it is scope creep.

4.6 Teacher / school layer (Phase 2, architected now)

Architect the data model so that a Class / Teacher entity can own a group of Students. Do not build teacher UI in Phase 1. But do not build Student UI in a way that makes teacher UI hard to add later. Specifically:

This avoids the classic mistake where consumer-first edtech has to be re-architected to sell to schools. We want the architecture ready, even if the UI waits.

---

5. Starky 2.0 — the upgrade

5.1 What Starky is today (as I understand it)

A chat-based AI tutor that answers questions when the student asks. Reactive. Session-by-session. No memory of the student between sessions. No awareness of what the student is working on.

5.2 What Starky 2.0 must be

The upgrade has five dimensions. All five must ship in the upgrade or Starky stays reactive.

**5.2.1 Smarter (pedagogy)**

**5.2.2 Persistent (memory)**

**5.2.3 Proactive (initiative)**

**5.2.4 Multimodal (capability)**

**5.2.5 Embedded (omnipresence)**

5.3 Emotional intelligence (the underrated layer)

Students — especially teenagers preparing for high-stakes exams — carry real stress. Starky must handle this without becoming a therapist. Specifically:

- If a student expresses stress, anxiety, or frustration, Starky acknowledges it first, solves second. Not the other way around.
- If a student expresses anything that suggests serious mental health risk (self-harm language, hopelessness, suicidal ideation), Starky:
1. Responds with care but does not play counsellor
2. Surfaces region-appropriate crisis resources (UAE: Dubai Foundation for Women and Children; Pakistan: Umang; India: iCall; UK: Childline)
3. Notifies the parent via a flagged dashboard event (within 1 hour, never silent)
4. Logs the event for platform safety review
- Starky does not discuss romance, sexuality, politics, or religion beyond what the curriculum requires. If asked, Starky politely redirects to the parent or teacher.
- Starky does not become a best friend substitute. If a student spends too much time chatting with Starky in ways unrelated to learning, Starky gently suggests other activities and, eventually, flags the pattern to parents.

This emotional safety layer is not optional. It is a legal, ethical, and reputational requirement. Any claim of "AI tutor for children" is a claim of duty of care.

5.4 The Starky prompt architecture

Starky is not one prompt. It is a layered system:

All four layers are composed at runtime and sent to Claude Sonnet (for serious teaching) or Claude Haiku (for lightweight chat, Brain Games explainers, in-game responses). Model selection is per-task, not per-session. Roughly 80% of calls should be Haiku (cost efficiency), 20% Sonnet (pedagogy-critical moments).

5.5 What Starky is NOT

---

6. Parent layer (the retention secret)

Every feature above fails commercially if the parent cannot see it working. The parent dashboard is therefore a core architecture concern, not an afterthought.

6.1 The weekly parent summary

Every Sunday morning, the parent receives a single email (and push notification, and WhatsApp if enabled) with a one-page summary:

Three sentences, one celebration, one concern, one action. No more. Parents who receive long reports stop reading after week three.

6.2 The curriculum impact view

A parent can open the app and see, within 3 seconds:
- How my child is doing this week vs. last week
- What my child is studying right now
- What my child is struggling with
- What Starky recommends next

No dashboards of metrics. No graphs that require interpretation. Sentences, status indicators, and one recommended action.

6.3 What parents cannot see

Parents trust the platform because it protects the child's private space too. Full visibility would feel surveillance-y and would lose older students immediately.

---

7. Data model (the spine)

The entire system hinges on a clean data model. Get this right and everything else is easier.

**Entities:**
- `User` — any human account (parent, student, teacher, admin)
- `Student` — a child's learning profile (may be linked to one User if self-managed, or to a Parent User)
- `Parent` — role on a User
- `Teacher` — role on a User
- `Class` — container owned by a Teacher, contains Students
- `Subject` — Cambridge Math, Edexcel English, etc.
- `Unit` — second-tier curriculum container
- `Lesson` — smallest learning unit
- `Attempt` — a Student's attempt at a Question or Mastery Check
- `BrainGamesSession` — one module session
- `BrainGamesScore` — per-Student, per-Module, time-series
- `StarkyMemory` — structured profile per Student, updated by Starky
- `StarkyConversation` — full chat log per Student (encrypted, parent-inaccessible by default)
- `SafetyEvent` — logged whenever safety layer fires
- `ParentEvent` — logged whenever parent is notified

**Key rules:**
- All Student data is scoped by `student_id`, never user_id
- Parent access to Student data is via an explicit `parent_student_link` join table — not implicit via user_id
- Every table has `created_at`, `updated_at`, `deleted_at` (soft delete)
- Every table has RLS policies from day one (see security brain in global ~/.claude/CLAUDE.md)
- No PII in application logs, ever

---

8. The build sequence

This is the order of work. Claude Code does not skip ahead. Each phase is reviewed by Khurram before the next begins.

**Phase 0 — Data model + Starky 2.0 foundation (Week 1–2)**
- Database migration for new entities
- RLS policies
- Starky base identity prompt + safety layer
- Memory storage structure

**Phase 1 — Platform Education Layer Tier 1 (Week 3–5)**
- Subjects / Units / Lessons schema and seed data for one exam board (Cambridge IGCSE Math as pilot)
- Lesson player UI (Explainer, Worked Example, Practice, Mastery Check, Review)
- Integration with existing 89k question bank
- Diagnostic Assessment (pilot: math only)
- Learning Path generator (pilot: one subject)

**Phase 2 — Brain Games MVP (Week 6–8)**
- Two modules to start: Memory Lab and Focus Training (highest immediate parent value)
- Master prompts for each
- Scoring engine (Cognitive Profile, Streaks, Mastery Badges)
- Starky explainers + curriculum bridges
- Integration with Lesson flow (cognitive gap detection → drill recommendation)

**Phase 3 — Parent Layer (Week 9–10)**
- Parent dashboard (3-second view)
- Weekly summary email
- WhatsApp integration (Dubai / Pakistan audience expectation)
- Notification preferences

**Phase 4 — Starky 2.0 full (Week 11–12)**
- Proactive initiative
- Multimodal (photo + voice)
- Curriculum Impact Score calculation
- Emotional safety layer full deployment

**Phase 5 — Remaining Brain Games modules (Week 13–14)**
- Pattern Recognition, Spatial Reasoning, Decision & Logic
- Full Learning Path coverage for Cambridge + Edexcel + FBISE + UAE MoE

**Phase 6 — Soft Launch (Week 15)**
- Limited rollout to existing paying users
- Telemetry, feedback, bug triage
- Retention and churn measurement

**Phase 7 — Public Launch (Week 16+)**

Total realistic timeline: 16 weeks with disciplined scope. Anything faster is a lie. Anything slower means scope bloat crept in.

---

9. What we are deliberately NOT building in this spec

Scope fences. These are real temptations that will waste time and must be declined unless Khurram explicitly reopens them later:

---

10. Risks and honest caveats

Things I want Khurram to think about before approving this.

1. **Cognitive Impact Score is a marketing claim.** We are saying Brain Games improves grades. We need to prove this with real data before we advertise it. Phase 6 must collect the data honestly, and we must be willing to NOT advertise this if the data does not support it. Overstating it damages the brand.

2. **Starky's memory layer creates data-protection obligations.** GCC, UK, and EU data rules apply the moment a child's learning profile is stored. Specifically: GDPR-K (for EU users), UAE Federal Decree-Law No. 45 of 2021, and forthcoming UK Online Safety Act implications. Legal review required before Phase 4 ships. I am not a lawyer.

3. **Starky's safety layer is the highest-risk component.** One mishandled crisis conversation becomes a front-page story. The safety layer needs real-world testing with child psychologists, not just engineering. Budget for this in Phase 0.

4. **The 16-week timeline assumes Khurram is the only decision-maker.** If schools, investors, or partners are looped in mid-build, timelines break. Protect decision-making velocity.

5. **Haiku quality drift.** If Anthropic updates Haiku in a way that changes Brain Games puzzle quality, we find out when parents complain. Implement a regression test: for each module, keep 20 "golden puzzles" and run them through the current model weekly. Alert on drift.

6. **Competitors will copy.** Once this ships, expect Cambridge, Pearson, or a US-based edtech to launch similar features within 12–18 months. The moat is speed to market, integration quality, and emerging-market fit — not any single feature.

7. **Brain Games may not actually improve grades.** The research on cognitive training transfer is contested. It is plausible that our Brain Games make children better at Brain Games but not at school. We must be honest with ourselves about this during Phase 6 data collection.

---

11. What I need from Khurram before Claude Code starts

Seven decisions, each of which shifts the spec:

1. Is the Phase 1 curriculum list correct? (Cambridge, Edexcel, FBISE, UAE MoE)
2. Is the 16-week timeline acceptable, or do we need to cut scope for faster launch?
3. Is $9.99/month still the right price point for this upgraded product, or does it move to $14.99 / $19.99?
4. Is Khurram comfortable with the safety-layer obligations (crisis detection, parent notification, professional review)?
5. Does the existing 89k question bank have the tags (subject, unit, difficulty, cognitive skill) needed for the Lesson tier, or do we need a tagging pass first?
6. Who is the first 50 users? We need real children using this in Phase 6 before public launch. Khurram's own children? Partner schools?
7. Do we want Brain Games to be included in the core $9.99 subscription, or sold as a $4.99/month add-on? (I recommend included — it is the retention moat.)

---

12. The global CLAUDE.md layer

This project inherits Khurram's global ~/.claude/CLAUDE.md which contains personal coding rules and the security brain. Both apply. Specifically:

---

13. Open questions for future iterations (not blockers)

None of these block Phase 1. Park them.

---

End of brief

This brief is complete at the architecture level. If Khurram approves it, the next document is a per-phase implementation brief that Claude Code receives, one phase at a time. Those briefs include specific files to create, specific schema, specific UI routes, specific test requirements.

This brief alone does not instruct Claude Code to build anything. It instructs Claude Code to wait for phase-level briefs.

**Signed:** Claude (architecture)
**Waiting for:** Khurram (approval, with or without changes)
**Date:** 2026-04-19

← Brainlab Kids — Strategic Platform BlueprintGGS Merge Build — Paste Into Claude Code →
Two years of working thought, indexed.
Ask me to present it in your conference room — WhatsApp +971 55 623 9111
Book Session →