Working notes for Hong Kong primary and secondary schools: planning, marking, guided practice, results data, data governance, and how to spend the AI Education Grant so there is something to show for it. Written to be lifted straight into a school plan.
Gemini, Gamma, Edpuzzle, Khanmigo, Perplexity, OpenAI Codex, Claude Code, Cursor and Base44 are not competing on the same track. Three lines separate them: whether the product treats you as a consumer, a developer or an institution; whether its terms let students in; and whether it bills per person or per school. Official September 2026 figures, plus the arithmetic for 100 staff.
One set of school usage assumptions applied to September 2026 official pricing from OpenAI, Anthropic, Google, xAI, Meta, Mistral and Cursor Composer, plus DeepSeek, Qwen, Kimi, GLM, MiniMax, Doubao, Hunyuan and ERNIE. Capability gaps are narrowing while price gaps widen, which means the real decision is not which model to use but who absorbs the volatility and what a generational change costs.
Base44 is Wix's vibe-coding app builder — describe what you need and you get a web app with a database, login and hosting, with no code written. This guide covers the September 2026 plans and allowances (Free, Starter US$16/month, Builder US$40, Pro US$80, Elite US$160, all annual), why message credits and integration credits behave differently, the age-13 minimum, and the full life of a panel-level utility from idea to handover.
The same "DeepSeek" means two entirely different privacy positions depending on whether you download the weights or call the API. This piece lays out the price, context window, licence and hardware floor for DeepSeek, Qwen, Kimi, GLM, MiniMax, Doubao, Hunyuan and ERNIE, then sorts school work into the routes a Hong Kong school can defend under the Personal Data (Privacy) Ordinance.
Claude Code comes with every paid Claude plan and not with Free. This guide sets out the September 2026 official prices (Pro US$20/month, Max from US$100, Team standard seats US$20–25 per user, Enterprise US$20 per seat plus usage), what it means that the five-hour rolling window is shared with your web chats, why the 18-and-over policy closes the student side completely, and a full walkthrough turning a timetable export into 60 individual staff timetables.
Cursor is Anysphere's AI code editor. As of September 2026 the plans are Hobby free, Pro US$20/month, Pro+ US$60, Ultra US$200, Teams US$40/user and Teams Premium US$120/user. This guide covers the two usage pools and what Composer 2.5 really costs, the Cursor Token Rate surcharge on Teams, what the 18-and-over terms mean for student accounts, and a full walkthrough of using it to teach debugging on the projector.
Edpuzzle is free for teachers, but the free Basic plan stores only 20 activities. This guide covers where to place checkpoints, what to check after auto-generated questions and Autograde, the Supported Assignments allowance trap, and the real cost structure behind Pro Teacher and Pro School, plus parental consent duties and the missing Traditional Chinese interface.
An AI agent is not a chatbot with better answers. It is a system that has been granted permission to act — it files a request, calls your school's data, and takes another step based on what comes back. This article follows one tool call end to end, explains what the MCP standard actually changes for a school, how per-call tool billing adds up, and the three risks that agency amplifies.
Gamma meters AI usage in credits, a free account gets 400 of them once at sign-up with no monthly refresh, and the terms of use require members to be at least 16. This guide walks through a controlled way to build a Secondary 3 subject-selection evening deck, works out where credits actually go, looks honestly at per-seat billing across 100 staff, and marks the place Gamma belongs in a school.
Gemini for Education is included at no extra cost in every Google Workspace for Education edition, but the Gemini sidebar inside Docs and Gmail needs the Google AI Pro for Education add-on at US$20 per user per month, restricted to users aged 18 and over. This guide separates the three layers, shows a controllable way to build differentiated worksheets, prices 100 licences honestly, and sets out what oversight and Traditional Chinese support really amount to.
Khan Academy's Khanmigo for Teachers is free to educators in more than 180 countries and territories and Hong Kong qualifies, but the paid student tutor at US$4 a month or US$44 a year is officially United States only. This guide separates the two product lines, states what Hong Kong can actually use, and shows how to land Socratic hints in the classroom.
In September 2026 the gap between OpenAI GPT-6 Astra and Google Gemini 3.8 Flash flagship input pricing is 13x, yet panel heads often cannot tell the summaries apart. Rather than a spec dump, this piece starts from real Hong Kong school tasks and walks through price, context, Chinese handling, access paths and billing traps.
A language model can cite a curriculum guideline number in perfect format that does not exist — a structural property, not one vendor's bug. Starting from everyday Hong Kong school scenes, this piece maps OWASP top risks and the NIST framework to hallucination consequences, what injection looks like, and the guardrails and audit arrangements a school platform must have.
Many schools assume a machine running an open model must beat cloud pricing. Run real usage through September 2026 published rates and the annual API bill is often too small for hardware to catch. This piece works both sides honestly, lists open models that actually fit a server room, and names the data-policy moments that tip the balance toward on-premises deployment.
Codex is OpenAI's agentic coding product — terminal CLI, IDE extension, cloud tasks and an iOS app — included with ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise. This guide covers September 2026 official prices, the credit rate card, GPT-5.6 versus GPT-6 Astra usage, why the free US K-12 teacher plan does not reach Hong Kong, and a full CLI walkthrough for a teacher-built classroom utility.
Every Perplexity answer carries its sources, which makes it strong for teaching verification. The Comet browser is free, Pro is US$20 a month or US$17 a month billed annually, and Enterprise Pro is US$40 per seat per month. This guide runs a Hong Kong Citizenship and Social Development project, supplies a pupil citation-checking sheet, and flags the age-13 clause and the absence of teacher monitoring.
Models in 2024 rushed to answer; models in 2026 reason internally first and decide their own thinking depth. This article explains chain-of-thought, reasoning effort and adaptive thinking in staffroom language, compares the four vendors' effort controls, and sets out what has genuinely changed for marking, maths and science — and what has not.
Starting from the minute a pupil types "just give me the answer" — what a school-based AI learning platform has to do. Curriculum alignment, a scaffolded tutor, live teacher intervention, four layers of guarding, role permissions, encrypted personal data, data ownership, swappable models, report data and on-site support.
The AI Education Grant (『智』啟學教) gives eligible schools up to HK$500,000 to spend by 31 August 2028, in return for three subjects across two grades each, six teaching exemplars, three demonstration lessons and three sharing sessions. Here is a budget split that funds people over licences, worked backwards from the interim report.
A procurement checklist you can paste straight into a tender. Exit cost, where the export button lives, the shape of the bill, whether the model can be swapped, and the boundaries around student data — with what a good answer sounds like and what should make you stop.
On-premises and cloud hosting each suit different schools, but neither should differ on who owns the data — it belongs to the school, can be exported in full free at any time, and can be reached at ai.school.edu.hk or your own subdomain. This piece helps Hong Kong schools choose by IT capacity, maintenance rhythm and contract terms.
A lesson plan, a slide skeleton and three tiers of worksheets in thirty minutes — using P5 General Studies (community planning) and S2 Geography (urban renewal), here is how those minutes should be spent, which part AI can take, and which part only the teacher can decide.
How a scaffolded AI tutor narrows support through signpost, handrail and one-step demo rather than ghostwriting; how teachers set a support ceiling and monitor live. Dialogue examples from P3 General Studies and Form 1 science show why "instant-answer" AI weakens learning.
Starting from a stack of thirty-two S2 English journals — how reliable AI marking really is across objective items, worked steps and open writing, how finely a rubric has to be written, the three gates teachers keep, and what feedback students will actually act on.
One learning objective can open three doors, each with its own SEN and non-Chinese-speaking adaptations. This article separates the part of differentiation worth delegating to AI from the part that breaks when you delegate it, and covers marking, assigning and tracking.
Why spaced repetition (the SM-2 algorithm) beats last-minute cramming, how AI generates flashcards from lesson texts and schedules reviews by each student's performance, and how teachers use review data to find the class's weakest concepts.
When students say they watched it, they are usually telling the truth, and playing a video to the end is not the same as processing it. This article turns one teaching video into subtitles, segment summaries, three check questions and a misconception list, then does the part that matters most, which is rebuilding the opening of the next lesson around what comes back.
Thirty-two pupils in one speaking lesson means under a minute of talking each. Starting from that arithmetic — what one round of AI oral practice actually looks like, how English and Putonghua differ in focus, where speech recognition genuinely fails, and how recordings are kept and reviewed.
Aligned with the new primary Science curriculum, this article shows how teachers turn a topic into a complete inquiry activity in 20 minutes with AI — question, hypothesis, variables, procedure, safety notes, record sheet and extension — with three worked examples and an inquiry quality checklist.
A mark sheet tells you who scored what; a Student Data Analysis System answers why, what the trend is and what to do next. Eight questions show how standard scores, percentiles, progress rates, multi-year tracking and teacher analysis are used — and how AI natural-language queries let every teacher ask.
Hong Kong teachers spend about 8–10 hours a week on paperwork. Six eAdmin scenarios show how AI drafts eNotice, eCircular, parent letters, meeting minutes, event proposals and iPortfolio comments — and why grounding in the school knowledge base and matching the school's voice decide whether a first draft is usable.
Every school runs Plan–Implement–Evaluate, but evaluation often relies on impressions. See how PIE KPIs link directly to platform data — AI usage, objective mastery, teacher training records, SDAS results — so planning has a baseline, implementation has tracking, evaluation has evidence, and grant interim and final reports export from the same cycle.
When primary pupils use AI, schools worry about three things — inappropriate content, personal data leaks and no oversight. This article lays out a workable architecture — input guard, system prompt, output review, audit & quotas — plus PII encryption, no-training terms, human review and PDPO mapping.
Generic AI has never read your school rules, pacing charts or notice formats. A school knowledge base uses RAG (retrieval-augmented generation) to search school documents before answering and attach source citations. This article covers how it works, what to upload, permission tiers and export requirements, and six practical scenarios for teaching and administration.
