香港中小學 AI 教育路線圖：四個階段，由「用過」到「造得出」An AI Education Roadmap for Hong Kong Schools: Four Stages, from Using AI to Building With It

香港的中小學不缺 AI 課程，缺的是次序。Hong Kong's schools do not lack AI lessons. They lack a sequence.同一個概念放在小二、中二與中五，要用三種完全不同的問法；同一個工具，今年教操作，兩年後可能已經不存在。次序錯了，學生會覺得 AI 很神奇但與自己無關；次序對了，他們會知道自己站在哪一步。 The same idea has to be asked three different ways in P.2, S.2 and S.5. The same tool taught this year may be gone in two. Get the order wrong and students find AI impressive but irrelevant to them. Get it right and they know which step they are standing on.

2026 年 6 月，教育局公布《中小學數字教育發展藍圖》^[15]，隨附《中小學人工智能素養學習架構》，把中小學分成四個學習階段，並把 AI 教育分成三條主線^[1][2]。這是香港第一份由初小到中六、橫跨所有科目的 AI 學習次序。In June 2026 the Education Bureau released the Blueprint for Digital Education Development in Primary and Secondary Schools^[15], with an AI Literacy Learning Framework attached. It divides schooling into four key stages and splits AI education into three strands^[1][2]. It is Hong Kong's first learning sequence for AI that runs from lower primary to S.6 and applies across every subject.

但框架寫的是「每個階段應該達到甚麼」。學校要的是「這個學期教甚麼」：用哪一個活動、哪一份功課、哪一個問題，去判斷學生是否真的學會。這也是 2026 年 2 月立法會上，教育界仍然在問的問題^[3]。文件已經有了，缺口由「有沒有框架」搬到了「怎樣落地」。But the framework says what each stage should reach. Schools need what to teach this term: which activity, which piece of work, which question tells you a student has actually learned it. That was still the question the sector was asking in the Legislative Council in February 2026^[3]. The document exists now. The gap has moved from having a framework to using one.

摘要Summary

1. 框架已經有了，次序未有。《中小學人工智能素養學習架構》寫清楚了四個階段各自要達到甚麼，但沒有寫九月第一堂做甚麼。The framework exists; the sequence does not. The AI Literacy Learning Framework states what each of the four stages should reach. It does not say what happens in the first lesson in September.
2. 分界線是兩個條件。學生有沒有親手做過那件事，以及做錯之後能不能復原；年齡只是代理指標。兩者決定一個概念是「可以教」還是「只可以講」。The dividing line is two conditions. Whether the student has done the task by hand, and whether an error can be undone. Age is only a proxy. Those two decide whether a concept can be taught or only described.
3. 初小與高小只看結果，中一才開始動手。前兩個階段不需要學生開任何工具：教師把 AI 的輸出帶進課室，學生判斷。每一段的產出都不同，不能提前借用作下一段的成果。Lower and upper primary see results; S.1 is where hands start. In the first two stages no student opens a tool: the teacher brings AI's output into the room and the students judge it. Each stage produces something different, and none can be borrowed as the next stage's outcome.
4. 學生缺少的是「非自動化不可」的經驗。香港 95% 的受訪學生用過 AI，但主要用途是找資料與做功課。他們從未做過一件重複到想交給機器的工作，所以聽不懂「為何要自動化」。What students lack is an experience that made automation necessary. 95% of the students surveyed in Hong Kong have used AI, mostly to find material and do homework. They have never done a job so repetitive that handing it to a machine was worth wanting, which is why the case for automation does not land.
5. agentic AI 與 harness 不應該是起點。它們同時要求最高的抽象度與最高的後果承受能力。但「誰批准、誰檢查、怎樣復原」這三個問題，中一已經可以問，也應該問。Agentic AI and the harness should not be the starting point. They demand the highest abstraction and the highest tolerance for consequences at the same time. But the three questions they raise, who approves, who checks and how you undo it, can be asked in S.1 and should be.
6. 「訓練 AI」有四層，學校通常只教到第一層。中一用圖片訓練分類器，是這一條路線合理的起點；微調語言模型與從頭預訓練是另外兩件事，不應該混為一談。There are four layers to training an AI, and schools usually reach only the first. Training an image classifier in S.1 is the reasonable starting point on this route. Fine-tuning a language model and pretraining one from scratch are different acts and should not be taught as one thing.

先做三十次，再自動化。重複到學生自己覺得煩，自動化才有意義；在那之前，教的是按鈕。Do it thirty times, then automate it. Automation means something once the repetition has become irritating. Before that, you are teaching buttons.

學校知道 AI 重要，卻卡在由哪一格開始Schools know AI matters. They are stuck on which square to start in

如果一間學校今天仍然覺得「不知道從何入手」，那是合理的，問題出在撥款與支援的形狀。三個具體的缺口同時存在：錢的用途偏向教師那一邊、人的訓練未到位、而「AI 素養」在 2026 年之前本來只覆蓋初中。If a school still feels it does not know where to begin, that is reasonable, and the cause sits in how the funding and support are shaped. Three specific gaps exist at once: the money leans towards the teaching side, the training has not caught up, and until 2026 AI literacy was formally a junior secondary matter only.

1. 錢的形狀，決定了課程的形狀1. The shape of the funding shapes the curriculum

「『智』啟學教」撥款計劃給每間公帑資助中小學一筆過 50 萬元，用途與承諾寫得很清楚：購置、訂閱或租用 AI 裝置與服務，以及資助學生參加提升 AI 素養的活動^[4]。撥款可用至 2027/28 學年完結，即 2028 年 8 月 31 日；中期報告在 2027 年 9 月前交，總結報告在 2028 年 9 月前交。The AI for Empowering Learning and Teaching Funding Programme gives every publicly funded primary and secondary school a one-off HK$500,000. Its permitted uses and the commitments are explicit: buying, subscribing to or leasing AI devices and services, and subsidising student activities that raise AI literacy^[4]. The money can be used until the end of 2027/28, that is 31 August 2028. The interim report is due by September 2027 and the final report by September 2028.

五項承諾之中，三項直接落在教師身上，一項落在學校層面，只有一項是關於學生學甚麼。撥款的目標本來就是「AI for ALL subjects」，即讓每個科目的教師都用得上 AI。但如果學校把那一項學生活動當成整個學生課程，得到的會是幾次活動，累積不成一條路線。Of the five commitments, three fall directly on teachers, one on the school, and only one on what students learn. The programme's stated aim is AI for all subjects, so that every subject teacher can use it. But a school that treats that single student commitment as its whole student curriculum will get a few activities, with no progression behind them.

表一：「『智』啟學教」的五項承諾，以及每一項主要落在誰身上。Table 1: the five commitments of the AI funding programme, and who each one mainly falls on.

承諾Commitment	數量Quantity	主要落在誰身上Mainly falls on
納入學校發展計劃及／或周年計劃Written into the school development plan and/or annual plan	1 項1 item	學校領導School leadership
在科目／課程範疇推行 AI 輔助教學，並發展教學例子AI-assisted teaching in subject or curriculum areas, with teaching examples	3 科 × 每科 2 級，共 6 個例子3 subjects × 2 levels, 6 examples	科主任與教師Panel heads and teachers
公開課或示範課Open or demonstration lessons	3 次3	教師Teachers
AI 輔助教學經驗分享會Sharing sessions on AI-assisted teaching	3 次3	教師Teachers
以提升學生 AI 素養與技能為重點的學生活動Student activities focused on AI literacy and skills	2 個2	學生Students

有一個容易踩的細節：撥款不可以直接用來資助教師修讀 AI 課程。教師培訓由另一條線負責，即每個三年 150 小時專業發展周期中，至少 30 小時的數字教育培訓^[5]。用撥款買培訓，會被視為不恰當運用；用 30 小時要求去安排培訓，才是正路。兩件事的錢與時間，不能互相替代。One easy mistake: the grant cannot be used to pay for teachers to attend AI courses. Teacher training sits on a different line, namely at least 30 hours of digital education training within each three-year, 150-hour continuing professional development cycle^[5]. Spending the grant on training is treated as improper use; planning training against the 30-hour requirement is the intended route. The money and the time are not interchangeable.

2. 使用率很高，熟練度不高2. High usage, low fluency

團結香港基金在 2025 年 7 月至 12 月訪問了 1,200 名校長、教師與學生，其中 833 名是學生。結果是教師使用率 91%、學生 95%，但教師自評 AI 熟練度（10 分滿分）只有 5 分，學生 6.5 分^[6]。使用與理解之間的距離，在校園裡同樣存在。Our Hong Kong Foundation surveyed 1,200 principals, teachers and students between July and December 2025, 833 of them students. It found 91% of teachers and 95% of students using AI tools, but teachers rated their own AI proficiency at 5 out of 10 and students at 6.5^[6]. The distance between using and understanding exists inside schools too.

同一份調查裡，只有 66% 的教師表示有教導學生在學習過程中使用 AI。按科目看，資訊及通訊科技科是 89%，語文與科學各約 70%，而數學科有 44% 的教師表示沒有教。AI 素養會不會教，很大程度上取決於學生坐在哪一間課室。In the same survey, only 66% of teachers said they teach students to use AI in their learning. By subject it was 89% in ICT, about 70% in languages and in science, while 44% of mathematics teachers said they do not. Whether a student meets AI literacy depends heavily on which classroom they happen to sit in.

3. 覆蓋範圍在一年之內由初中擴到全校3. In one year, the coverage moved from junior secondary to the whole school

在框架公布之前，香港的 AI 素養課程基本上只覆蓋初中；高小的增潤編程單元教的是算法思維，與 AI 的關係是間接的。這是團結香港基金在 2026 年 1 月提出的觀察，也是它建議「由初小開始、按學習階段制定漸進框架」的理由^[6]。五個月後，框架真的由初小開始了。Before the framework, AI literacy in Hong Kong was essentially a junior secondary matter. The upper primary enriched coding units taught algorithmic thinking, which relates to AI only indirectly. That was Our Hong Kong Foundation's observation in January 2026, and the reason it recommended starting in lower primary with a progressive framework^[6]. Five months later, the framework did start in lower primary.

政策方向是對的，但它把一個新問題交到小學手上：一批從未教過 AI 的班主任與常識科教師，下個學年要決定小二做甚麼活動。他們需要的是一條可以按學期執行的次序。The policy direction is right, but it hands primary schools a new problem: class teachers and general studies teachers who have never taught AI must decide, next school year, what a P.2 activity looks like. What they need is an order they can run a term at a time.

三份框架，同一個次序Three frameworks, one order

三份獨立製作的框架，用完全不同的名詞，卻排出同一個次序：先讓學生經驗與判斷，再理解運作，再動手造，最後才談放手與塑造。這是這個學科本身的形狀。Three independently written frameworks, using entirely different vocabulary, arrive at the same order: let students experience and judge first, then understand how it works, then build, and only then talk about delegation and shaping. That is the shape of the subject.

香港的《中小學人工智能素養學習架構》用三條主線貫穿四個學習階段：在資訊及通訊科技相關科目「學習 AI」、在各科與跨學科範疇「應用 AI」、以及「愛護地使用 AI 的態度」，後者包含倫理與道德^[2]。文件明言採用「螺旋式進展」，並以學生的認知發展為依據：初小着重感知與體驗，高小着重理解與培養興趣，初中着重邏輯建構，高中着重應用與價值創新。Hong Kong's AI Literacy Learning Framework runs three strands through four key stages: learning AI in ICT-related subjects, applying AI across subjects and cross-disciplinary areas, and an attitude of using AI with care, which carries the ethics^[2]. The document states that it uses a spiral progression grounded in students' cognitive development: lower primary emphasises perception and experience, upper primary understanding and interest, junior secondary logical construction, and senior secondary application and value innovation.

聯合國教科文組織 2024 年的《學生人工智能能力框架》用四個維度、十二項能力，配三個進階層次：理解、應用、創造^[7]。同一份文件給了一個冷水：截至 2022 年，全球只有 15 個國家把 AI 學習目標寫進國家課程。UNESCO's 2024 AI Competency Framework for Students uses four dimensions and twelve competencies across three progression levels: understand, apply and create^[7]. The same document offers a sobering figure: by 2022, only 15 countries had put AI learning objectives into their national curricula.

經濟合作與發展組織與歐盟在 2026 年 6 月發表的 AILit 框架，把學習分成四個依次出現的領域：接觸 AI、與 AI 共創、管理 AI、塑造 AI。文件說得很直白：共創與管理可以並行，而「塑造 AI」是最後一個領域，需要額外的支架與時間才發展得來^[8]。這份框架也會成為 PISA 2029 媒體與人工智能素養評估的基礎。The AILit framework, published by the OECD and the European Union in June 2026, divides learning into four domains that appear in sequence: engage with AI, create with AI, manage AI and shape AI. It says plainly that creating and managing can run in parallel, and that shaping AI is the culminating domain, which may need additional scaffolding and time to develop^[8]. The framework will also underpin the PISA 2029 media and AI literacy assessment.

表二：三份框架的結構對照。名詞不同，終點相近。Table 2: how the three frameworks are built. Different words, similar destinations.

框架Framework	怎樣分階段How it stages learning	終點是甚麼What the end looks like
香港（2026）Hong Kong (2026)	四個學習階段 × 三條主線Four key stages × three strands	建構 AI 輔助工具，並判斷其對社會的影響Build AI assistive tools and judge their impact on society
聯合國教科文組織（2024）UNESCO (2024)	三個層次：理解、應用、創造Three levels: understand, apply, create	成為 AI 的共同創造者與負責任公民Become AI co-creators and responsible citizens
OECD 與歐盟（2026）OECD and EU (2026)	四個領域，依次出現Four domains, in sequence	改善 AI 系統，使之反映人的價值Improve AI systems so they reflect human values

三者唯一明顯的分歧，是倫理放在哪裡：香港把它放進每一階段，聯合國把它列為一個維度，AILit 把它放進態度。這個分歧對學校的影響不大，因為三者都要求學生在同一件事上做判斷。The one clear disagreement is where ethics sits: Hong Kong puts it inside every stage, UNESCO makes it a dimension, and AILit folds it into attitudes. It matters little to a school, because all three ask students to make a judgement about the same act.

對學校最實用的讀法是：用香港那份對齊要求與家長溝通，用另外兩份檢查自己的次序有沒有跳步。三份文件在同一個位置都留白了：它們寫「達到甚麼」，不寫「用哪一個活動達到」。那一段是學校的專業判斷。The most useful reading for a school is this: use the Hong Kong document to align with requirements and to explain the plan to parents, and use the other two to check that your own sequence has not skipped a step. All three leave the same space blank: they say what to reach, not which activity gets you there. That part is the school's professional judgement.

兒童缺少的是自動化的理由What children lack is a reason to automate

我們習慣假設孩子「由細用到大」，所以自然懂。數據不支持這個假設。更重要的問題在別處：他們從未有過一個理由，去想象一件工作交給機器。We assume children have grown up with AI and therefore understand it. The data does not support that. The more important problem lies elsewhere: they have never had a reason to imagine handing a job to a machine.

同一份調查問學生用 AI 做甚麼，答案很集中：搜集資料 35%、做功課與專題報告 29%、整理課本重點 18%、生成練習題 7%、課外探索 6%^[6]。這五項全部是「找一個答案」或「起草一份東西」。沒有一項是流程：沒有先後次序、沒有例外情況、沒有要交給下一個人的步驟。學生由此建立的模型是「AI 是一個快一點的搜尋器」。這個模型本身說得通，只是 agentic AI 假設的是另一個模型。The same survey asked students what they use AI for, and the answers cluster tightly: finding material 35%, homework and projects 29%, summarising textbook points 18%, generating practice questions 7%, exploring outside the syllabus 6%^[6]. All five are either finding an answer or drafting something. None is a process: no sequence, no exceptions, no step that hands work to the next person. The model students build is that AI is a faster search box. That model holds up, but agentic AI assumes a different one.

成人學 agentic AI，靠的是一份積壓：他們做過那件事二百次，知道它在哪裡出錯，所以寫得出規格。一個中二學生沒有這份積壓。他從未填過三十份同一款表格，也從未把幾十張圖片分類。所以「讓 AI 幫你做」對他來說只是一個示範。這是經驗問題。Adults learn agentic AI from a backlog. They have done the task two hundred times, they know where it breaks, and so they can write the specification. A Form 2 student has no backlog. They have never filled in thirty copies of the same form, or sorted a few dozen images. So "let the AI do it for you" is a demonstration to them. The obstacle is experience.

一個警告：無限制的接觸會取代思考A warning: unfettered access substitutes for thinking

2024 年的一項實地實驗，把近一千名高中數學學生分成兩組，一組用一般聊天介面，一組用加上了學習護欄的版本。使用期間，兩組成績都上升，一般版本升 48%，護欄版本升 127%。但當工具被拿走之後，用一般版本的學生成績比從未用過的學生低 17%；加了護欄的一組，負面影響大致消失^[9]。研究者用的詞是「拐杖」：學生在練習時把思考外判了。A 2024 field experiment gave almost a thousand high school mathematics students access to generative AI tutors, one resembling a standard chat interface and one with prompts designed to safeguard learning. While the tools were available, grades rose: 48% with the standard version and 127% with the safeguarded one. But once access was removed, students who had used the standard version performed 17% worse than students who had never had access, while the safeguards largely removed the negative effect^[9]. The researchers' word for it was a crutch: during practice, the students had outsourced the thinking.

這個結果對課程設計的意義很具體：同一件工具，可以令表現變好而能力變差。所以「早一點給學生用」「用多一點」要成為一條課程，前提是設計上強迫他把判斷做回來。That result says something concrete about curriculum design: the same tool can improve performance while lowering ability. So "give students access earlier" and "use it more" only become a curriculum when the design forces the judgement back onto the student.

看不見的機制才是障礙The barrier is an invisible mechanism

另一邊的證據同樣重要。一項 2019 年的研究讓兒童用身體動作訓練一個手勢辨識系統，並把機器學習的兩個環節打開給他們看：標註資料，以及評估結果。研究發現，兒童能夠理解基本的機器學習概念，前提是兩個環節都要打開；只打開其中一個，理解就建立不起來。他們還能將這個理解帶到日常生活中^[10]。The evidence on the other side matters just as much. A 2019 study had children train a gesture-recognition system with their bodies while uncovering two steps of machine learning: labelling the data, and evaluating the result. It found that children could understand basic machine learning concepts, but only when both steps were uncovered; one alone was not enough. They also carried that understanding into everyday contexts^[10].

兩項研究放在一起，得到一條很實用的界線：決定一個概念難不難的，是機制可不可以被看見，以及學生有沒有親手做過那兩個動作；年齡只是代理指標。看不見的機制，講十次都是名詞；看得見的機制，小學生也看得明，分別在於誰的手在操作。Put the two studies together and a practical line appears. What makes a concept hard is whether the mechanism can be seen, and whether the student has done the two acts by hand. Age is only a proxy. An invisible mechanism stays a vocabulary word no matter how often it is explained. A visible one is legible even to a primary student; the difference is whose hands are on it.

三十次原則The thirty-times rule

所以課程需要的是一段刻意安排的「手工階段」。我們在課堂上用一個簡單的做法，稱為三十次原則：在把任何工作交給 AI 之前，學生先用最笨的方法做三十次。What the curriculum needs is a deliberate manual phase. In our classes we use a simple practice we call the thirty-times rule: before any task is handed to AI, the student does it thirty times the slow way.

三十是一個夠長的時間：長到規律看得見，長到厭倦變得真實。改三十份同類的短答、把六十張照片分類、核對二十條引文、翻譯三十句句子，都是這一類工作。做完之後，學生手上有三樣東西，而這三樣正好是一個代理所需的前提：一份規格（他知道步驟是甚麼）、一個錯誤測試（他知道怎樣算做錯）、和一個理由（他知道為何不想再做）。Thirty is simply long enough for the pattern to become visible and for the boredom to become real. Marking thirty short answers against a rubric, sorting sixty photographs, checking twenty citations, translating thirty sentences: these are the tasks. Afterwards the student holds three things, and those three are exactly what an agent presupposes: a specification, because they know the steps; an error test, because they know what wrong looks like; and a reason, because they know why they do not want to do it again.

這也是「為何需要 agentic AI」這個問題在中小學唯一講得通的答案。因為學生親手做過那件煩事。沒有那件煩事，代理只是一個魔術；有了那件煩事，代理是一個解法。It is also the only answer to "why do we need agentic AI" that works in a school. Because the student has done the tedious thing by hand. Without the tedium, an agent is a magic trick. With it, an agent is a solution.

三十次原則是我們的教學做法：「先做手工、再自動化」這個方向有研究支持，「三十」這個數字來自課堂經驗。學校可以按科目調整，要保留的是那段手工階段，數字可以改。The thirty-times rule is our teaching practice. The direction, manual work before automation, is supported by the evidence; the number thirty comes from classroom experience. A school can adjust it by subject. What must survive is the manual phase; the number can change.

五個詞，五種難度：由提示詞到 harnessFive terms, five kinds of difficulty: from prompt to harness

「AI 素養」這個詞在 2020 年才被整理成一組可討論的能力與設計考量^[14]，此後每份框架都加上自己的分層。香港、聯合國與 OECD 的分法不同，但都指向同一件事：AI 素養是一組能力。學校文件裡最常混用的五個詞，真正的分別是「你要指定甚麼」。The term AI literacy was only consolidated into a set of discussable competencies and design considerations in 2020^[14], and every framework since has added its own layers. Hong Kong, UNESCO and the OECD divide it differently but point at the same thing: AI literacy is a set of capabilities. The five terms schools most often use interchangeably differ in what you have to specify.

五個詞排成一條線：由指定答案的形狀，到指定來源、指定例子、指定目標與邊界，最後把邊界寫成規則。The five terms form a line: from the shape of an answer, to its source, to examples of behaviour, to the goal and the boundary, and finally to the boundary written as rules.

提示詞（prompt）Prompt

你用文字指定答案的形狀：題目、角色、格式、限制。你指定的是輸出，行為留給模型。You use words to specify the shape of an answer: the task, the role, the format, the limits. What you specify is the output; the behaviour is left to the model.

常被混淆Often confused with「AI 素養」本身。提示詞只是其中一個動作。AI literacy itself. A prompt is one move inside it.

最早適合Earliest stage中一。之前只判斷教師帶來的輸出，不自己開工具。S.1. Before that, students judge output a teacher brings in; they do not open a tool themselves.

知識庫（RAG）Retrieval (RAG)

你指定答案的來源：先在指定文件裡找，再根據找到的段落作答，並附上出處。它改變的是答案的依據，模型本身不變。You specify where the answer may come from: search the given documents first, answer from the passages found, and attach the source. What changes is the basis of the answer; the model stays as it is.

常被混淆Often confused with「訓練」。知識庫沒有訓練任何東西，它只是把文件放到模型面前。Training. A knowledge base trains nothing; it puts documents in front of the model.

最早適合Earliest stage中二至中三。前提是學生已有「資料有出處」的習慣。S.2 to S.3, once the student already expects a source for a claim.

微調（fine-tuning）Fine-tuning

你指定行為的例子：給模型一批「這樣問、這樣答」的示範，讓它模仿語氣、格式或判斷方式。它改變的是模型的傾向。You specify examples of behaviour: give the model a set of question-and-answer demonstrations so it imitates a tone, a format or a way of deciding. What changes is the model's tendency.

常被混淆Often confused with中一用圖片訓練分類器。那是訓練，但訓練對象是一個較小的分類模型。The S.1 image classifier. That is training, but of a much smaller classification model.

最早適合Earliest stage中五以後，而且只在有明確評估準則時才做。S.5 and beyond, and only with an explicit evaluation rubric.

智能代理（agentic AI）Agentic AI

你指定目標與邊界，中間的步驟由模型自己決定，並自行取用工具。業界的分法是：固定流程叫 workflow，由模型自己決定流程才叫 agent^[11]。You specify the goal and the boundary; the model decides the steps and uses tools itself. The industry distinction is that a fixed path is a workflow, and only a system that directs its own process is an agent^[11].

常被混淆Often confused with「自動化」。自動化按你寫好的步驟走；代理會自己選步驟，所以出錯的方式也不同。Automation. Automation follows the steps you wrote; an agent chooses them, which is why it fails differently.

最早適合Earliest stage中四至中五，而且由「限制它做甚麼」開始。S.4 to S.5, and it starts from what the agent may not do.

Harness（外圍支架）Harness

你把邊界寫成規則，交由系統在每次執行時自動套用：它可以呼叫哪些工具、要誰批准、最多試幾次、記憶保留多久。模型不變，改變的是它被允許做甚麼。You write the boundary as rules the system applies on every run: which tools it may call, who must approve, how many attempts it gets, how long memory is kept. The model does not change. What changes is what it is allowed to do.

常被混淆Often confused with「提示詞」。提示詞是請求，harness 是規則；學生可以改前者，只有設計者可以改後者。A prompt. A prompt is a request; a harness is a rule. A student can change the first; only the designer changes the second.

最早適合Earliest stage概念在中一可以教，名稱與實作留到中五以後。The idea is reachable in S.1; the name and the implementation can wait until S.5.

這個分法解釋了兩件事。第一，代理之所以難，是因為它要求學生由「講答案」跳到「講邊界」，中間跳過了「講來源」與「講例子」兩級。第二，知識庫應該排在代理之前，因為它先建立一個習慣：答案要有依據。有了這個習慣，「行動要有授權」才講得通。This split explains two things. First, agents are hard because they ask the student to jump from specifying an answer to specifying a boundary, skipping the two rungs in between: specifying a source and specifying an example. Second, retrieval belongs before agents because it establishes a habit: an answer needs a basis. Once that habit exists, "an action needs authorisation" makes sense.

Harness 這個名稱在 2025 年之後才在業界流行，但概念在香港的框架裡已經出現了：小六的學習重點包括「配置合適的設定規範人工智能的使用」，以及「監察和檢查人工智能的判斷和生成結果」^[12]。換句話說，「設定界線」這個動作可以在小學的討論裡出現；到中一，學生開始自己操作，這件事才變成一個要回答的問題；知道它叫甚麼、怎樣寫成程式，留到中五。概念可以早，名稱不必早。The name harness only became common in the industry after 2025, but the idea is already in Hong Kong's framework: a P.6 learning focus asks students to configure appropriate settings to regulate the use of AI, and to monitor and check its judgements and outputs^[12]. In other words, the act of setting a limit can appear in a primary discussion; in S.1, once students begin operating AI themselves, it becomes a question they have to answer; knowing what it is called and how to write it as code waits for S.5. The idea can come early. The name does not have to.

分階段路線圖：每一階段教甚麼、不教甚麼、怎樣知道學會了The roadmap, stage by stage: what to teach, what to leave out, and how you know it landed

四個階段直接對應香港框架的四個學習階段。每一段都有一個「離開時應該做得到」的測試，那個測試最實用，因為它把「教過了」變成「學會了」。教育局同時公布了一批應用情境示例，示範同一條主線如何放進不同科目與不同階段，並提醒 AI 只應作為支架，隨學生掌握程度逐步減少^[16]。我們在其中加了一條自己的界線：中一之前，學生不開工具。初小與高小看到的，是 AI 的結果；判斷一張圖片哪裡不合理，不需要密碼。The four stages map directly onto the four key stages in Hong Kong's framework. Each ends with a test of what a student should be able to do on leaving it, and that test is the most useful part, because it turns coverage into evidence. The Bureau also published a set of application scenarios, showing how one strand enters different subjects and stages, and noting that AI should serve only as scaffolding, reduced as students gain mastery^[16]. We add one line of our own: before S.1, students do not open a tool. What lower and upper primary see is AI's output, and pointing out what is wrong in an image needs no login.

初小：教「分辨」Lower primary: teach recognition

框架對初小的要求很克制：知道 AI 在日常生活中的應用、知道它會生成內容，而且明言不建議長時間使用電子裝置，使用時要有教師或家長在旁^[2]。這個克制是有理由的：這個年紀的學生仍在建立「甚麼是機器、甚麼是人」的分界。我們的做法比框架再收緊一步：初小的 AI 課由教師把結果帶進來，學生不開工具、不開帳戶。The framework is restrained about lower primary. Students should know where AI appears in daily life and that it generates content, and the document explicitly discourages long sessions on devices, with a teacher or parent present^[2]. The restraint has a reason: at this age students are still building the boundary between what is a machine and what is a person. Our practice goes one step tighter than the framework: in lower primary the teacher brings the output in, and students open no tool and no account.

所以初小最值得做的兩個活動，是分類與找錯。分類是香港框架裡已有的做法：把圖卡分成「有 AI」與「沒有 AI」，例如語音助手與計算機、推薦影片與電子時鐘。找錯則是教師帶來一張 AI 生成的圖片，讓他們指出不合理的地方：沒有門的音樂室、課室裡的滑梯。這一個動作是往後所有批判能力的起點，因為它教的是「我可以查」。相反，這個階段最不值得做的是解釋模型原理，因為學生沒有可以掛住這個解釋的經驗。So the two activities worth doing here are sorting and fault-finding. Sorting is already in Hong Kong's framework: cards divided into AI and not-AI, a voice assistant against a calculator, a recommendation feed against a digital clock. Fault-finding brings in an AI-generated image and asks what does not make sense: a music room with no door, a slide inside a classroom. That single move is the root of every later critical skill, because what it teaches is that I can check. What is not worth doing at this stage is explaining how a model works, because there is no experience for the explanation to attach to.

離開這一階段的測試：學生能拿起一件由 AI 生成的東西，指出一個具體位置，並說出一個理由。理由要具體：「這隻手有六隻手指」。The test on leaving this stage: a student can pick up something AI-generated, point to a specific place, and give a reason. The reason should be specific: "this hand has six fingers".

高小：看一次完整的訓練與測試Upper primary: watch one full cycle of training and testing

高小的框架內容明顯加厚：輸入、處理、輸出，數據、算法、算力三大要素，提示詞的運用，以及用數據訓練 AI 並分析結果^[2]。教育局的小學課程文件把這一段拆得更細：小四判斷哪一句提示較好；小五用圖片訓練分類器並測試，然後刻意找一批有偏差的數據再訓練一次，比較結果；小六自行設計提示，並分析偏見的成因^[12]。這份文件把訓練放在小五、小六，由學生自己操作。我們的建議是把操作權留在教師手上：一堂課用一部電腦走完那個循環，學生負責預測與記錄。Upper primary is where the framework thickens: input, process and output; the three elements of data, algorithms and computing power; the use of prompts; and training AI with data and analysing the result^[2]. The Bureau's primary curriculum document breaks it down further. P.4 judges which of two prompts is better. P.5 trains an image classifier, tests it, then deliberately gathers a skewed set of images, retrains and compares. P.6 designs prompts independently and analyses where bias comes from^[12]. That document puts training in P.5 and P.6, in the student's own hands. Our recommendation keeps the operating in the teacher's: one computer runs the cycle, and the students predict and record.

這一段裝下了三件日後很難補救的事：數據會影響結果、結果要用測試判斷、以及「我改一個變數，看看會怎樣」的習慣。教師示範一次「收集二十張圖片、訓練、測試、換一批偏斜的數據、再比較」，學生在每一步之前猜結果，在每一步之後解釋差異。看見整個循環，勝過親手按完每一個按鈕。This stage fits in three things that are hard to make up later: data changes the result, a result has to be judged by a test, and the habit of changing one variable to see what happens. The teacher runs the cycle once, collecting twenty images, training, testing, swapping in a skewed set and comparing, while students predict each result before it appears and explain each difference after. Seeing the whole cycle is worth more than pressing every button in it.

同一個階段最不應該做的，是提早教「微調語言模型」。它需要評估準則、足夠的示範數據與判斷品質的能力，而這三樣都要再等幾年。離開這一階段的測試：學生能預測一次訓練的結果，並在它失敗時說出一個可能的原因，例如「我給它的圖片全部都是同一種角度」。What should not be pushed into this stage is fine-tuning a language model. It needs an evaluation rubric, enough demonstration data and the ability to judge quality, and all three are years away. The test on leaving: the student can predict what a training run will do, and can name a plausible cause when it fails, for instance that every image given to it came from the same angle.

初中：第一次自己動手，由「來源」開始Junior secondary: the first hands-on stage, starting with sources

初中框架的關鍵詞是機器學習流程、自然語言處理、偏見與幻覺，並明確要求學生「用不同數據訓練 AI 系統，分析數據如何影響結果」^[2]。同時，框架把「vibe coding」列為資訊及通訊科技科的 AI 應用例子之一。這一點值得學校注意：自然語言編程已經在課程文件裡面，屬於課堂內容。中一也是這條路線第一次由學生自己開工具：先寫提示，再學核對，然後才碰知識庫。The junior secondary keywords are the machine learning workflow, natural language processing, bias and hallucination, with an explicit requirement that students train AI systems on different data and analyse how the data changes the result^[2]. The framework also names vibe coding as an example of an AI application in ICT. That is worth a school's attention: natural-language programming sits inside the curriculum document. S.1 is also where a student first opens a tool on this route: write a prompt, learn to check it, and only then touch a knowledge base.

這一段最值得做的一個專題，是造一個會引用出處的助手，然後把它弄壞。學生先把一份試卷或一組筆記放進知識庫，讓助手只根據這些材料作答並附上段落；接著問一個材料裡沒有的問題，看它會怎樣答；再決定要它說「不知道」，還是要它去查另一個來源。整個過程教的是一個習慣：答案要有依據，而沒依據的時候要說沒有。這比任何一次「提示詞技巧」課都接近真實工作。The project most worth running here is to build an assistant that cites its sources and then break it. Students put a past paper or a set of notes into a knowledge base and require the assistant to answer only from that material, with the paragraph attached. Then they ask something the material does not contain, see what it does, and decide whether it should say it does not know or go and check another source. The habit being taught is simple: an answer needs a basis, and when there is no basis you say so. That is closer to real work than any lesson on prompt tricks.

離開這一階段的測試：學生能造出一個助手，指出它在哪裡失效，並提出一個檢查方法。例如「它答課外題時會編，所以我們加了一句：材料裡沒有就答不知道」。The test on leaving: the student can build the assistant, say where it fails, and propose a check. For instance, it invents an answer to a question outside the material, so we added a rule that it must say it does not know.

高中：教「授權」，因為這一步開始有後果Senior secondary: teach authorisation, because this is where consequences start

高中框架的要求是「為社區開發創新的 AI 方案」，並理解 AI 對社會、行業與公共服務的影響^[2]。這一句話把學生由使用者推向設計者。設計者要回答的問題不再是「怎樣問得更好」，而是「它可以做甚麼、不可以做甚麼、誰批准、出錯怎樣復原」。The senior secondary framework asks students to develop innovative AI solutions that contribute to the community, and to understand AI's effect on society, industries and public services^[2]. That sentence moves the student from user to designer. The designer's question is no longer how to ask better. It is what it may do, what it may not do, who approves, and how an error is undone.

四個問題剛好可以寫成四行規格，而這四行就是我們要求高中學生交的東西：可以做（列出工具與資料範圍）、不可以做（列出禁止動作）、誰批准（列出需要人手確認的步驟）、怎樣復原（列出紀錄與撤銷方法）。寫得出這四行，學生才算明白甚麼是代理；寫不出，他只是用過一個代理。Those four questions fit on four lines, and those four lines are what we ask senior students to hand in: what it may do, listing the tools and the data in scope; what it may not do; who approves, listing the steps that need a human; and how to undo, listing the log and the rollback. A student who can write those four lines understands what an agent is. A student who cannot has only used one.

離開這一階段的測試：學生能為一個代理寫出這四行，並且指出一個「它技術上做得到，但不應該做」的任務。那個判斷，就是這一整條路線的終點。The test on leaving: the student can write those four lines for an agent and name one task it could technically perform but should not. That judgement is where the whole route ends.

同一個概念，四個階段的問法One concept, four ways of asking

「太難」通常是因為問法沒有換。同一個詞，在四個階段可以是完全不同的問題。A concept is rarely too hard in itself. The question has usually not been changed. The same word becomes a different question at each stage.

表三：四組概念在四個階段的問法。左邊是概念，右邊四欄是同一個概念的四種難度。Table 3: four concepts asked at four stages. The left column is the concept; the four to its right are four levels of the same question.

概念Concept	初小Lower primary	高小Upper primary	初中Junior secondary	高中Senior secondary
AI 出錯AI is wrong	指出圖片不合理的地方Point at what is wrong in a picture	用第二個來源核對Check it against a second source	寫出評估準則並解釋Write the criteria and explain them	設計檢查步驟，估算錯誤成本Design the check; estimate the cost of error
訓練Training	留到高小Left to upper primary	用二十張圖片訓練分類器Train a classifier on twenty images	用不同數據訓練，比較結果Train on different data and compare	用示範微調，並定下評估準則Fine-tune on examples, with a rubric
私隱Privacy	不輸入自己與別人的資料Enter no personal data, yours or anyone's	說出哪些資料會被收集Say which data gets collected	分析權限與保留期Analyse permissions and retention	設計資料流向與刪除安排Design the data flow and the deletion
自動化Automation	指出生活裡的自動門Point at the automatic door	說出自動化的步驟次序Say the steps in order	把流程寫成步驟與例外Write the process as steps and exceptions	決定哪一步可以放手Decide which step can be released

agentic AI 為何排在最後Why agentic AI comes last

短答：因為它同時是抽象度最高、後果最大的一格。把兩個維度放在一起，一個概念該排在路線的哪個位置，就一目了然。The short answer: it sits in the one square that is both the most abstract and the most consequential. Put those two dimensions together and the position of a concept on the route becomes visible.

表四：用抽象度與後果兩個維度定位概念。位置是判斷；重點是代理落在最難的一格。Table 4: placing concepts by abstraction and consequence. The positions are judgements; the point is that agents land in the hardest square.

抽象度Abstraction	後果低：錯了一句話Low consequence: a wrong sentence	後果高：錯了一個動作High consequence: a wrong action
低：看得見、可示範Low: visible, demonstrable	對話提示詞圖像生成知識庫ChatPromptsImage generationRetrieval	固定流程自動化機械人自動評分Fixed workflowsRobotsAutomatic marking
高：看不見、要想象High: invisible, imagined	微調偏見分析模型比較Fine-tuningBias analysisModel comparison	代理自行決定步驟Harness 設計Agent-directed stepsHarness design

這一格最難，因為它同時要求兩樣在校園裡最稀缺的東西：學生要能想象一個自己看不見的流程，而且要有能力承受一個做錯了的動作。前者可以靠課程補，後者要靠制度補。That square is the hardest because it demands the two scarcest things in a school at once: the student must be able to imagine a process they cannot see, and must be able to live with an action that went wrong. The first can be taught. The second has to be built into the system.

業界的定義可以幫學校把話說準。固定流程、由人寫好每一步的系統叫 workflow；由模型自己決定步驟、自己取用工具的系統才叫 agent^[11]。學校現時很多稱為「agent」的活動，其實是 workflow：學生設定好觸發條件與步驟，系統按次序執行。這沒有問題，甚至是對的教法；問題只是叫錯了名。叫錯名的代價，是學生以為自己見過代理，於是跳過了最需要學的那一段。The industry definition helps a school say the right thing. A system whose steps are written in advance is a workflow. Only a system that directs its own steps and picks up its own tools is an agent^[11]. Much of what schools currently call an agent activity is in fact a workflow: students set the trigger and the steps, and the system runs them in order. That is fine, and it is the right way to teach the machinery. The only mistake is the name. The cost of the wrong name is that students think they have met an agent, and skip the part they most need.

真正的風險是「無邊界地示範」。如果一堂課只展示代理可以做甚麼，沒有展示它被禁止做甚麼，學生只學到能力。這種示範比不教更難逆轉，因為它建立了一個印象：只要做得到，就可以做。The real risk is demonstrating agents without boundaries. A lesson that shows only what an agent can do, and never what it is forbidden to do, teaches capability only. That is harder to reverse than no lesson at all, because it leaves the impression that anything possible is permitted.

三個問題，中一已經可以問Three questions an S.1 class can already ask

代理這個概念的核心是授權。而授權是校園裡最熟悉的事。用一個具體例子：班長負責收通告。如果他可以代同學簽名，這是權限過大；如果他只可以收、不可以簽，這是有邊界；如果收錯了可以退回重收，這是可復原。三個問題一問，學生立刻明白「代理」是甚麼，因為他們每天都活在同一個結構裡。The core of an agent is authorisation, and authorisation is the most familiar thing in a school. Take a concrete case: a class monitor collects permission slips. If the monitor may sign on a classmate's behalf, the permission is too broad. If the monitor may collect but not sign, there is a boundary. If a slip collected wrongly can be returned and redone, the action is reversible. Ask those three questions and students understand what an agent is, because they already live inside that structure every day.

這三個問題在香港的框架裡有對應的位置：小六要學「配置合適的設定規範人工智能的使用」，以及「監察和檢查人工智能的判斷和生成結果」^[12]。官方文件已經把「邊界」放進了小學課程，只是沒有用 harness 這個詞。我們把它放在中一講，因為那時學生已經開始自己操作，授權才不是一個抽象名詞。Those three questions already have a home in Hong Kong's framework. P.6 asks students to configure appropriate settings to regulate the use of AI, and to monitor and check its judgements and outputs^[12]. The official document has already put a boundary into the primary curriculum without using the word harness. We teach it in S.1, because by then students have begun operating AI themselves, and authorisation stops being an abstract noun.

所以準確的說法是：概念要早，實作要遲。中一教「誰批准、誰檢查、怎樣復原」，中五才教「怎樣把這三件事寫成規則」。中間那幾年，是靠一次次的專題把三個問題問熟。至於 harness 這個詞，學生在中四之後自然會遇到，到時他會發現自己早就懂那個道理，只是未學那個名。So the accurate rule is: the idea early, the implementation late. S.1 teaches who approves, who checks and how to undo. S.5 teaches how to write those three things as rules. The years in between are spent asking the three questions often enough that they become habit. Students will meet the word harness after S.5, and by then they will find they already know the idea and simply did not know its name.

學校本身就是一個 harness。時間表、點名、家長同意書、權限、留堂紀錄，全部都是「在執行之前先設定的規則」。學生一直活在裡面，只是沒有人告訴他這就是設計。把這一層指出來，比任何一個示範都有效。A school is itself a harness. The timetable, the register, the consent form, permissions, the detention record are all rules set before anything runs. Students have lived inside one all along and nobody told them it was a design. Naming that layer does more than any demonstration.

所以代理的位置在中間：起點是經驗，終點是判斷。代理只是把兩者連起來的一段路。把這一段路提前，學生會看見一個很聰明的東西；把它放在對的位置，學生會知道自己為甚麼需要它。So the agent sits in the middle: the start is experience, the finish is judgement. An agent is the stretch of road that joins them. Move that stretch forward and students see something clever. Put it where it belongs and they understand why they need it.

「訓練一個 AI」其實有四層，學校通常只教到第一層There are four layers to "training an AI", and schools usually reach the first

「訓練」是中英文裡最容易被含混過去的一個詞。它至少指四件不同的事，而四件事的難度、成本與風險相差很遠。Training is the most overloaded word in the field. It refers to at least four different acts, and their difficulty, cost and risk are far apart.

表五：「訓練」的四層。由第一層到第四層，改變的東西越來越深，學校的控制力越來越低。Table 5: four layers of training. From the first to the fourth, what changes goes deeper and the school's control goes down.

層Layer	改變了甚麼What changes	校內對應活動Classroom version	最早階段Earliest stage
提示Prompting	沒有改變模型，只改變這一次的輸入Nothing in the model; only the input this time	改一句提示，比較兩個結果Change one prompt and compare two results	中一S.1
檢索Retrieval	模型不變，答案的依據變了The model is unchanged; the basis of the answer changes	把一份試卷放進知識庫，要求附出處Put a past paper in a knowledge base; require the source	中二S.2
微調Fine-tuning	模型的行為傾向變了，需要評估準則The model's behaviour shifts; a rubric is now required	用三十組示範，訓練一種固定的答題格式Thirty demonstrations to fix an answer format	中五S.5
預訓練Pretraining	整個模型由數據生成，學校不可控The whole model is produced from data; outside a school's control	只講概念，不做實作Explain the idea; do not attempt it	不適用Not applicable

中小學做得到的，是首兩層；第三層要有評估準則才做；第四層只適合講，不適合做。但在學校的簡報裡，四件事常常都被叫做「訓練 AI」。這會產生兩個實際問題。Schools can do the first two layers. The third needs a rubric before it is worth doing. The fourth is a topic to explain, not an activity to run. Yet in school presentations all four are routinely called training an AI, which creates two practical problems.

第一個問題是對外說錯了能力。如果一間學校說「我們的學生訓練了 AI」，而實際做的是寫提示，家長與辦學團體會得到一個過高的印象；到學生升學面試被追問細節時，落差會出現。第二個問題是對內教錯了重點：微調的失敗很細微，要看評估集才知道；訓練分類器的失敗很明顯，換一張圖片就看得見。對初學者來說，看得見的失敗才有教學價值。The first is that it overstates what students can do. If a school says its students trained an AI when they wrote prompts, parents and sponsoring bodies form too high an impression, and the gap shows when a student is questioned in an admissions interview. The second is that it misdirects the teaching. Fine-tuning fails subtly, in ways only an evaluation set reveals; a classifier fails visibly, on the next image you show it. For a beginner, a visible failure is where the teaching is.

所以有一個簡單的排序規則：要教某一層，先教它下面那一層。想教微調，先讓學生做過檢索；想教檢索，先讓學生會寫提示並且會核對；想教提示，先讓學生做過評估，知道兩個答案哪個較好和為甚麼。這個規則的好處是，它不需要新的資源，只需要改課程的次序。So there is a simple ordering rule: to teach a layer, teach the one beneath it first. Before fine-tuning, students should have done retrieval. Before retrieval, they should be able to write a prompt and check the answer. Before prompting, they should have practised judging which of two answers is better and saying why. The rule needs no new resources. It only changes the order.

他們畢業之前，這些東西會全部換掉Before they graduate, most of this will be replaced

一個今年讀小一的學生，大學畢業大約是 2042 年。今天任何一個工具的名稱，大概都不會出現在他的履歷上。所以課程要分兩層來看：哪些是會過時的例子，哪些判斷仍然成立。A child starting P.1 this year graduates from university around 2042. No tool name in use today is likely to appear on their CV. A curriculum therefore has two layers: the examples, which expire, and the judgement, which still holds.

連僱主也不假設穩定。世界經濟論壇 2025 年的調查中，僱主預期到 2030 年有 39% 的核心技能會改變，比 2023 年的 44% 略為回落^[13]。這個數字同時說明兩件事：變化仍然很大，但沒有一般人想象的那麼線性。用一個預測去設計六年課程，風險很高；用一個判斷去設計，風險低得多。Even employers do not assume stability. In the World Economic Forum's 2025 survey, employers expected 39% of workers' core skills to change by 2030, slightly down from 44% in 2023^[13]. The figure says two things at once: change is still large, and it is less linear than usually imagined. Designing a six-year curriculum around one forecast is risky. Designing it around a judgement is much less so.

AILit 框架自己也是這樣定位的：它說要建立「持久的基礎」，聚焦於「超越特定工具與潮流」的能力^[8]。三份框架在這件事上一致，因為這是課程設計的基本算術：工具的壽命比學制短。The AILit framework positions itself the same way. It sets out to build a durable foundation, focused on competences that transcend specific tools or trends^[8]. All three frameworks agree here, because this is the arithmetic of curriculum design: tools have shorter lives than a school career.

有一個寫法可以直接用：把工具放在例子裡，不要放在目標裡。目標寫「學生能判斷一段生成的內容是否適合某個場合，並修改它」；例子寫「用一個圖像生成工具做一張海報」。工具換了，目標不動，教案只需要換例子。這一條同時解決了採購鎖定的問題：如果課程目標沒有綁定任何一個產品，學校就可以換供應商而不必重寫課程。A writing rule follows directly: put the tool in the example, never in the objective. The objective is that students can judge whether generated material suits a given occasion and revise it. The example is making a poster with an image generator. When the tool changes, the objective stands and only the example is swapped. The same rule solves procurement lock-in: if no objective names a product, a school can change supplier without rewriting its curriculum.

評估也要守同一條線。香港框架對評估的要求很清楚：不應該只有紙筆測驗與背誦，要看應用能力、實作能力與創新思維，並且重視過程而非結果，題目要貼近真實情境^[2]。用這個標準檢查一份現成的評估：如果它考的是「按下哪個按鈕」，它會跟工具一齊過期；如果它考的是「你會怎樣判斷」，它仍然成立。Assessment has to hold the same line. Hong Kong's framework is explicit: assessment should not be confined to pen-and-paper tests or rote recall, should cover application, practice and innovative thinking, should value process over outcome, and should be anchored in authentic situations^[2]. Run any existing assessment through that filter. If it tests which button to press, it expires with the tool. If it tests how you would judge, it stays.

一間學校可以怎樣開始：一個學年，四個時段How a school can start: one year, four windows

不要由一份三年計劃開始。由一個科目、兩個級別、一個可以拿出來給人看的成果開始。這也正好符合撥款的形狀，因為撥款要求的是六個教學例子、三次公開課、三次分享會與兩個學生活動。學校層面的使用原則另有官方指引^[17]。Do not start with a three-year plan. Start with one subject, two levels and one outcome you can put in front of someone. That also matches the shape of the funding, which asks for six teaching examples, three open lessons, three sharing sessions and two student activities. School-level principles for use are set out in a separate official guide^[17].

人手：一個人做四件事，不如四個人各做一件Staffing: four people doing one thing each beats one person doing four

最小可行的組合是一位統籌教師，加上每個參與科目一位科主任。統籌教師負責次序、紀錄與對外溝通；科主任負責那一個科目的問法與活動。不建議成立一個「AI 小組」把四件事交給同一位老師，因為那個人很快就會變成行政協調員。The minimum workable structure is one coordinating teacher plus one panel head for each participating subject. The coordinator owns the sequence, the record and the communication. The panel head owns the way the concept is asked in that subject and the activity that carries it. Do not form an AI team that hands all four jobs to one teacher; that person becomes an administrator very quickly.

採購：五條問題，問完才簽Procurement: five questions before you sign

撥款可以買軟件、硬件、平台、學與教資源與校本方案^[4]。買之前問五條問題，每一條都與課程能不能持續有關。The grant can buy software, hardware, platforms, learning and teaching resources and school-based solutions^[4]. Before buying, ask five questions. Each is about whether the curriculum can continue.

• 學生需要個人帳戶嗎？如果需要，資料存放在哪裡、保存多久、誰可以查看^[18]。Does it need student accounts? If so, where the data sits, how long it is kept, and who can see it^[18].
• 學校可以匯出學生的作品與資料嗎？不能匯出，學校就沒有評估證據，也沒有辦法在停用時取回成果。Can the school export student work and data? Without export there is no assessment evidence, and nothing can be recovered when the tool is dropped.
• 如果這個工具兩年後停運，課程還跑得動嗎？如果跑不動，代表課程目標綁定了產品。If the tool closed in two years, would the course still run? If not, the objective is tied to a product.
• 教師可以修改內容與提示嗎？只有供應商可以改，就代表課程每年要等一次更新。Can teachers change the content and the prompts? If only the vendor can, the curriculum waits for a release every year.
• 有沒有不用這個工具的替代活動？有替代，學校才有議價能力，也才有停用時的退路。Is there an equivalent activity without this tool? An alternative is what gives a school leverage, and a way out.

最後有兩個提醒。第一，兩個學生活動是撥款的下限；只做兩個活動，學生帶走的是兩個回憶。第二，平台只提供情境，次序仍然要學校自己排，判斷仍然要教師自己教。Two cautions to close. First, the two student activities are a funding floor; run only those and students take away two memories. Second, the platform supplies a situation. The sequence is still the school's to set, and the judgement is still the teacher's to teach.

學校不需要一次做完四個階段。它需要知道今年做哪一格，以及下一格是甚麼。知道這兩件事，就不會出現「買了工具但沒有人用」的年度，也不會出現「活動很熱鬧但能力沒有累積」的三年。No school has to do all four stages at once. It has to know which square it is doing this year, and what the next one is. With those two answers, a school avoids the year in which the tools were bought and nobody used them, and the three years in which the activities were lively and nothing accumulated.

五個常見誤解，每一個都會令課程走錯方向Five common misconceptions, each of which bends a curriculum the wrong way

這五句在學校工作坊裡最常聽到。每一句都有一半是真的，而錯的那一半會決定課程的次序。These five sentences come up most often in our school workshops. Each is half true, and the wrong half is what decides the order of a curriculum.

• 「他們由細用到大，所以識。」用得多與看得見機制是兩件事。香港 95% 的受訪學生用過 AI，但用途集中在找資料與做功課；教師自評熟練度 5 分，學生 6.5 分。真正要知道的是：他能不能說出一個答案為甚麼錯。"They have grown up with it, so they understand it." Using a lot and seeing the mechanism are different things. 95% of the students surveyed have used AI, mostly to find material and do homework, and self-rated proficiency is 5 out of 10 for teachers and 6.5 for students. The question that matters is whether a student can say why an answer is wrong.
• 「直接教 agent，因為那是未來。」它同時是抽象度最高與後果最大的一格。先教固定流程的 workflow，學生會明白步驟與例外；再教代理，他才會明白「由它決定」多了甚麼、少了甚麼控制。"Teach agents straight away, because that is the future." Agents sit in the square that is both most abstract and most consequential. Teach a fixed workflow first and students learn steps and exceptions. Teach agents after that and they can see what letting the model decide adds, and what control it removes.
• 「AI 素養等於提示詞技巧。」提示詞是三條主線裡「應用 AI」的一個動作。素養還包括機制（數據、算法、算力）、評估（準則與核對）、與倫理（私隱、偏見、學術誠信）。只教提示，等於只教了三條主線裡的一格。"AI literacy means knowing how to prompt." A prompt is one move inside one of three strands. Literacy also covers mechanism (data, algorithms, computing power), evaluation (criteria and checking) and ethics (privacy, bias, academic integrity). Teach only prompting and you have taught one square of one strand.
• 「學編程已經沒有用。」香港框架在初中階段仍然要求學生應用算法、處理數據、以編程解決問題，並把 vibe coding 列為其中一種應用。自然語言編程減少了語法記憶，拆解、除錯與判斷照樣要做。"Learning to program is pointless now." Hong Kong's framework still asks junior secondary students to apply algorithms, handle data and solve problems by coding, and it names vibe coding as one such application. Natural-language programming removes syntax memorisation. Decomposition, debugging and judgement still have to be done.
• 「評估就是考操作。」框架明確反對以紙筆測驗與背誦為主，要求看應用、實作與創新，重視過程而非結果，並且要有真實情境。考操作還有一個實際問題：工具一換，成績就無法比較。"Assessment means testing the operation." The framework explicitly rejects a diet of pen-and-paper tests and rote recall, and asks for application, practice and innovation, with process valued over outcome and authentic situations. Testing the operation has a practical flaw as well: when the tool changes, the results stop being comparable.

我們怎樣教這一套How we teach this

我們的課程針對中學生，位置大約在這條路線的中段與後段。Our courses are built for secondary students, which places them in the middle and later part of this route.

自建 AI 學習助手對應的是本文中一至中三那一段：學生在視覺化畫布上，把一份試卷變成知識庫，加上分流器，接上幾個專科助手，造出一個會先查自己的文件再作答的助手。整個過程沒有寫過一行語法，但學生必須回答三個問題：文件夠不夠、答案有沒有出處、答錯了怎樣知道。Build Your Own AI Tutor corresponds to the S.1 to S.3 part of this article. On a visual canvas, students turn a past paper into a knowledge base, add a router and connect subject assistants, so the result checks its own documents before answering. No syntax is written, but students have to answer three questions: are the documents enough, does the answer carry a source, and how would you know it is wrong.

自然語言編程與人工智能開發教的是拆解與除錯。六小時由零到一個上線的應用，重點是「我怎樣知道自己要的是甚麼」。這正好是代理之前必須具備的能力：說得出步驟，才寫得出規格。Natural-Language Programming and AI Development teaches decomposition and debugging. Six hours take a student from nothing to a live app, and the point is knowing what you want. That is precisely the ability an agent presupposes: if you can state the steps, you can write the specification.

月球車探測機械人補上另一件事：後果。機械人的錯是一個撞上去的動作。學生在這門課上第一次需要問「如果它做錯了，我怎樣停」。這個問題日後會變成代理的權限設計。The lunar rover programme adds the thing the others cannot: consequence. A robot's error is an impact. It is the first time many students have to ask how to stop it if it goes wrong, and that question later becomes the design of an agent's permissions.

學校層面，我們的到校課程與「智啟學教」配合方案可以承接撥款要求的兩個學生活動，也可以為教師做一次兩小時的簡報，把這條路線變成校內共同的語言。至於初小，我們不派導師入班：那個年紀最合適的人是班主任，我們提供的是活動設計與教師簡報。At school level, our on-campus programmes and the AI funding package can carry the two student activities the grant asks for, and a two-hour briefing can turn the sequence into a shared vocabulary for teachers. For lower primary we do not send an instructor into the room. At that age the class teacher is the right person, and what we supply is the activity design and the teacher briefing.

如果家長的問題是「先學 AI 素養還是先學編程」，〈AI 素養還是編程？〉處理的是同一個次序，只是由家長的角度問。如果問題是「Vibe Coding 到底是甚麼」，〈甚麼是 Vibe Coding？〉由定義講起。If the question is whether to learn AI literacy or coding first, 〈AI Literacy or Coding?〉 takes the same order from a parent's point of view. If the question is what vibe coding actually is, 〈What Is Vibe Coding?〉 starts from the definition.

學校最常問的問題Questions schools ask most

香港中小學的 AI 教育應該由哪一年級開始？Which year should AI education start in?

由初小開始，起點是「分辨」。香港框架要求初小學生知道 AI 在日常生活中的應用、知道它會生成內容，並在教師或家長指導下使用^[2]。可執行的活動是分類（哪些東西有 AI）與找錯（這張圖哪裡不合理）。學生自己開工具，由中一開始。Lower primary, with recognition as the starting point. Hong Kong's framework asks lower primary students to know where AI appears in daily life, to know that it generates content, and to use it under a teacher's or parent's guidance^[2]. The workable activities are sorting familiar objects by whether they use AI, and finding what is wrong in a generated image. Students open a tool themselves from S.1.

初小學生可以用生成式 AI 嗎？Can lower primary students use generative AI?

不建議，初小的 AI 課不需要學生碰工具。框架明言在初小階段不建議長時間使用電子裝置，使用時要有教師或家長在旁，並且要建立負責的使用習慣^[2]。實際做法是由教師操作、學生觀察與判斷，學生不開個人帳戶，不輸入自己或同學的資料。Not in an AI lesson. The framework states that prolonged use of devices is not recommended at this stage and that use should happen under a teacher's or parent's guidance, building responsible habits^[2]. In practice the teacher operates and the students observe and judge. Students do not open individual accounts and do not enter their own or anyone else's data.

agentic AI 是不是太難，要等大學？Is agentic AI too hard, and should it wait for university?

概念不用等，實作要等。代理的核心是授權，而授權可以用校園情境教：誰批准、誰檢查、怎樣復原。這三個問題中一已經可以問，香港框架在小六亦要求學生「配置合適的設定規範人工智能的使用」^[12]。把三個問題寫成規則並實作，才留到中五以後。The idea does not have to wait; the implementation does. The core of an agent is authorisation, and authorisation can be taught with a school situation: who approves, who checks, how it is undone. An S.1 class can ask all three, and Hong Kong's framework already asks P.6 students to configure settings that regulate the use of AI^[12]. Writing those three answers as rules and building them can wait for S.5.

中學生可以「訓練 AI」嗎？Can secondary students train an AI?

可以，但要分清四層。用圖片訓練一個分類器是中一最好的入門，因為失敗看得見。檢索（把文件放進知識庫）中二可以開始。微調語言模型要有評估準則與示範數據，通常中五以後。從頭預訓練不適合學校做，只適合講。Yes, but the four layers have to be separated. Training an image classifier is the best entry point in S.1 because failure is visible. Retrieval, putting documents into a knowledge base, can start in S.2. Fine-tuning a language model needs a rubric and demonstration data, so usually S.5 and later. Pretraining from scratch is a topic to explain, not an activity for a school.

學校沒有 AI 科主任，可以由誰教？A school has no AI panel head. Who teaches it?

由科主任教本科的問法，由一位統籌教師負責次序。框架本身就不把 AI 素養放在單一科目：學習 AI 在資訊及通訊科技相關科目，應用 AI 橫跨各科與跨學科範疇^[2]。這也符合撥款要求的三個科目各兩個級別。與其找一位 AI 專家，不如找三位願意改一道題目的科主任。The panel head teaches the way the concept is asked in their own subject, and one coordinating teacher owns the sequence. The framework does not put AI literacy in a single subject: learning AI sits in ICT-related subjects, while applying it runs across subjects and cross-disciplinary areas^[2]. That also matches the funding requirement of three subjects at two levels each. Three panel heads willing to change one question each are worth more than one AI expert.

怎樣評估 AI 素養？How should AI literacy be assessed?

看過程、看情境、看判斷。框架要求評估不應局限於紙筆測驗與背誦，要看應用與實作，重視過程而非結果，題目要貼近真實情境，並且要涵蓋倫理與批判思考^[2]。可操作的版本是：給一個真實任務，收三樣證據：學生的作品、學生對結果的解釋、教師的觀察。避免考「按下哪個按鈕」。Look at process, situation and judgement. The framework asks that assessment not be confined to pen-and-paper tests or rote recall, that it cover application and practice, that it value process over outcome, and that it be anchored in authentic situations and include ethics and critical thinking^[2]. The workable version is one real task and three pieces of evidence: the artefact, the student's explanation of it, and the teacher's observation. Avoid testing which button to press.

「智啟學教」撥款可以用在這些活動嗎？Can the AI funding programme pay for this?

可以買軟件、硬件、平台、學與教資源與校本方案，也可以資助學生參加提升 AI 素養的活動；不可以資助教師修讀 AI 課程^[4]。撥款可用至 2027/28 學年完結，即 2028 年 8 月 31 日。教師培訓走另一條線：每個三年 150 小時專業發展周期中至少 30 小時的數字教育培訓^[5]。It can buy software, hardware, platforms, learning and teaching resources and school-based solutions, and it can subsidise student activities that raise AI literacy. It cannot subsidise teachers attending AI courses^[4]. The money runs to the end of 2027/28, that is 31 August 2028. Teacher training runs on a separate line: at least 30 hours of digital education within each three-year, 150-hour professional development cycle^[5].

這條路線會不會兩年後就過時？Will this roadmap be out of date in two years?

例子換得快，次序留得久。用一條測試檢查每一堂：如果這個工具明天消失，學生還剩下甚麼？把工具寫在例子裡，不要寫在目標裡。學科內容會隨課程文件更新，但拆解、核對、判斷與授權這四件事，在模型換代之後仍然成立。The examples age quickly. The order lasts. Test every lesson with one question: if this tool disappeared tomorrow, what would the student still have? Put the tool in the example, not in the objective. Subject content will change with the curriculum documents, but decomposition, checking, judgement and authorisation still hold when a model is replaced.