B2C content packs (Kwi Quick-Starts)¶
Static, segment-targeted content packs that give every B2C learner something concrete to do in one click. Kwi recommends from the learner's segment pack on /learn, upgrading the recommendation from "an intent" to real content. Generated once with LLMs, human-reviewed once, seeded as data. No per-user generation.
Why this exists¶
- Who asked: Bhanu (owner), 2026-07-16, after watching the shipped Kwi recommendation still point at bare intents: "we should offer some tailored content already from our side... hot courses, coding puzzles."
- User pain: a class 6 learner lands on
/learn, Kwi suggests "Understand a topic", and the kid still has to invent what to do. Grades 1-10 have zero prebuilt inventory; Study Plan serves only 11-12 JEE/NEET. - Cost of not doing: the persona layer stays cosmetic. Kids without homework in hand bounce.
- Validated or guess: guess, and this ships BEFORE the parent survey by explicit owner override (static LLM generation is cheap enough to test in production). Kill/expand per segment by data, not opinion.
- How we know it worked:
kwi_action_clickedand pack activity completion per segment; a segment whose pack nobody finishes gets replaced.
Segments¶
Derived entirely from data onboarding already captures: class_level (1-12) and exam_family (JEE/NEET/null). Cut points follow NEP 2020 stages and age research (logic stumpers from ~8, real programming from ~7-9 with formal exposure at grade 6, boards pressure at 9-10).
| Segment | Classes | Framing |
|---|---|---|
| foundational | 1-3 | audio-first, picture-led, 3-minute wins |
| preparatory | 4-5 | computational thinking enters, streak mechanics |
| middle | 6-8 | real coding starts, collection motivation |
| secondary | 9-10 | boards-aware, nothing babyish |
| senior | 11-12, no exam intent | subject depth (PCM) + projects |
| jee (overlay) | any + exam_family=JEE | real PYQ bank only |
| neet (overlay) | any + exam_family=NEET | PYQ chemistry + reviewed LLM bio/physics |
Exam overlays take precedence over the class segment on /learn.
Pack contents (4 per segment)¶
Classes 1-3: Number Ninja (daily 5-question picture math) · Puzzle Zoo (pattern/sequencing puzzles, pre-coding logic) · Word Sparks (phonics riddles + picture vocabulary) · Story Time with Kwi (micro-story + 2 comprehension taps)
Classes 4-5: Mental Math Tricks (speed arithmetic ladders) · Logic Lab (grid puzzles, odd-one-out, mini Sudoku) · Code Blocks (move-the-robot sequencing, no syntax) · Curious Why (science curiosity + home mini experiments)
Classes 6-8: Python Basics track (first syntax, 3-minute challenges) · Math Olympiad Lite (tiered IMO/NMTC-style) · Brain Gym (memory training, mnemonics, speed drills) · Science Sprints (chapter-aligned rapid quizzes + diagram labeling)
Classes 9-10: Board PYQ Drills (class-10 pattern by chapter) · Web Dev Challenges (HTML/CSS/JS micro-builds + Python continuation) · Aptitude & Reasoning (NTSE-style) · English Edge (spoken prompts, debate topics)
Classes 11-12 (no exam): Maths Concept Drills (calculus, algebra, chapter-wise) · Physics Problem Sets (mechanics, electricity, numericals) · Chemistry Rapid Rounds (organic reactions, physical chem numericals) · Python Mini Projects (data, automation) (Economics/accountancy stream dives and CUET prep were considered and rejected: owner call, negligible usage expected.)
JEE overlay: Weightage-ranked PYQ drills per topic · Formula flash decks (feeds future spaced review) · 10-minute topic rapid mocks · Common Traps (most-missed PYQs, worked explanations)
NEET overlay: Chemistry PYQ drills · Biology fact sprints + mnemonic decks · Diagram labeling · Assertion-Reason format practice
Sourcing rules¶
- Exam overlays use the real PYQ bank (6,493 topic-tagged questions; JEE complete, NEET chemistry). Real past questions are the trust anchor; LLM fills only NEET biology/physics, flagged as practice items, never presented as past questions.
- Everything else is LLM-generated once, against the pack spec, then human-reviewed before seeding. Child-facing content review is non-negotiable (quality + POCSO).
- All AI-generated items carry the platform's AI-content label.
- v1 grading is client-side and the activity payload ships answer keys to the browser. Acceptable while packs are zero-stakes static practice; the moment pack results feed the learner model (adaptive stage 1), grading must move server-side so mastery signal can't be gamed.
Non-goals¶
- No per-user or on-demand generation. Packs are static data.
- No adaptivity in v1. Same pack for every learner in a segment; the learner model (see B2C adaptive learning) personalizes later.
- No new pedagogy engine: activities reuse existing runners (practice session player, quiz formats) wherever possible.
- No marketplace/creator flow.
Metrics¶
- North star: % of segment-active learners completing ≥1 pack activity per week.
- Guardrail: chat depth and practice completion must not drop.
- Per-segment
kwi_action_clicked(surface: pack) and activity completion drive kill/expand decisions.
Relationship to Kwi¶
Kwi fronts the pack: the /learn recommended action draws from the learner's segment pack, celebration fires on completion, and the streak mechanic counts pack activity as practice.