The activity-based approach organizes learning around three levels of activity: activities of learning and understanding, in which students perceive and attend to the language and structure of the text; activities of applying and practicing, in which they internalize and use what they have grasped in new but structured contexts; and activities of transferring and creating, in which they evaluate, extend and produce meaning beyond the text [1]. The Reading and Thinking section of the textbooks was constructed to give this progression a concrete home, and its position at the head of each unit signals that the text is the unit’s principal source of language, structure and meaning.

Read as a design specification rather than as a description of lesson stages, the three levels define a cognitive gradient along which the questions of a lesson can be arranged. Questions that ask students to locate, sequence or classify information occupy the first level. Questions that ask them to explain a relationship, infer an unstated attitude, or apply a pattern to new material occupy the second. Questions that ask them to judge the author’s stance, weigh competing readings, or propose an alternative occupy the third. The gradient is homologous with the movement from remembering and understanding toward analyzing, evaluating and creating in revised taxonomies of educational objectives [8].

Reading comprehension research describes understanding as the construction of a coherent mental representation in which information from the text is integrated with the reader’s prior knowledge under the constraints of the text’s own organization [9]. Comprehension is therefore not the accumulation of extracted facts but the assembly of a structure, and the structure is precisely what a well-ordered sequence of questions can help a reader to build. A reader who is asked what happened, then why it happened, then what it shows about the character asking and being asked these questions, is being supported in constructing a situation model rather than in compiling a list [10].

The scaffolded character of this support connects to the notion of a zone of proximal development, in which a learner can accomplish with assistance what is currently beyond unassisted reach, provided that the assistance is calibrated to what the learner can already do [11]. A question chain operationalizes this calibration discursively. Each question establishes a piece of common ground, and the next question is answerable because of it. Difficulty is thus raised incrementally rather than abruptly, and the class is carried into territory it could not have entered directly.

This is what distinguishes a chain from a list. A list is a set of questions that happen to concern the same text; a chain is a set of questions in which the answer to one becomes a resource for answering the next. The distinction has consequences for design. In a list, each item can be evaluated independently, and a weak item can be dropped without loss. In a chain, removing a link severs a dependency, and the later questions become unanswerable or arbitrary. Research on classroom interaction has shown that the structure and functional quality of a teacher’s questioning shapes both the amount and the cognitive quality of student talk [12].

Systematic reviews of artificial intelligence in education have documented a persistent gap between the capability of the technology and its pedagogical integration [13]. GenAI renews the question because a single tool now spans functions that were previously distributed across several technologies: generating text at different levels of difficulty, producing questions, simulating an interlocutor, and commenting on written work [14].

The division of labour adopted in this paper follows from these observations and can be stated compactly: the machine proposes, the teacher determines. GenAI is treated throughout as a generator of candidate questions, and the teacher as the architect of the chain. The machine’s advantage is volume and speed under specification; the teacher’s advantage is judgement about evidence, about learners and about sequence. Section 4 converts this principle into six operational strategies, and Table 1 summarizes how it distributes across the three layers of the framework.

Table 1.The three layers of the framework and the distribution of work between GenAI and the teacher.

The function of the first layer is to make the text available as an object of thought. Questions here direct attention to what the text says, how it is organized, and how its parts refer to one another. Typical functions include locating and sequencing information, identifying the referents of pronouns and nominalizations, grouping details under categories, and recognizing the text’s organizational pattern. Because these questions are answerable by inspection and are also the easiest to formulate, they are the layer at which teachers most often stop, and the layer at which GenAI’s contribution is most nearly complete: a model can produce such questions in large numbers and at several levels of granularity within seconds.

The risk at this layer is abundance without discrimination. Because the questions are easy to generate, they are easy to generate too many of, and a lesson can consume its available time at a layer that requires the least of students. The teacher’s task here is therefore subtractive rather than additive: the relevant judgement is not whether a generated question is acceptable, but whether it is necessary, and specifically whether anything later in the chain depends on it. Questions that establish ground nothing else stands on should not be asked.

The second layer asks students to work with what the first layer has established. Its functions include explaining cause and effect, comparing and contrasting elements within and across texts, applying a linguistic or rhetorical pattern to new material, and inferring attitudes, intentions or implied meanings that the text supports but does not state. These functions demand that the reader hold several pieces of textual information in relation simultaneously and reason across them, which is where the cognitive distance from the first layer lies.

GenAI’s contribution at this layer is considerable but requires more supervision than at the first. Models are fluent at producing prompts of the form that ask students to explain why something is so, to compare two elements, or to infer a speaker’s attitude, and they can supply several alternative phrasings of the same prompt so that a teacher can choose the one best suited to the class’s proficiency. The corresponding risk is unwarranted inference. An inference question is only as good as the evidence available to support it, and a model may propose an inference whose textual basis is thin, partial or absent. Every second-layer question must therefore be checked against the specific words in the text that make it answerable.

The third layer asks students to move beyond the text. Its functions include evaluating the author’s stance or the adequacy of the text’s treatment of its subject, weighing competing interpretations against one another, transferring the text’s insights to a new situation, and producing a reasoned position of one’s own. These functions are generative: their answers cannot be verified by inspection, and different defensible answers may coexist. They are also, in most reading-and-thinking lessons, the functions that carry the developmental purpose of the section.

GenAI’s contribution at this layer is distinctive but must be handled carefully. A model can supply a range of interpretations of a text, including some that a teacher would not have considered, and it can argue a counter-position on demand, which makes it useful for constructing evaluative questions that present students with a genuine choice. The risk is that the model’s interpretations may be plausible-sounding recombinations of commonplaces, and that exposure to them may flatten rather than widen the interpretive range of a class. For this reason the culminating question of a lesson, the one that defines what the class has been building towards, should be formulated by the teacher rather than adopted from the model. Section 4 develops this as the sixth strategy.

The most common way to prompt a model for reading questions is to supply the text and ask for comprehension questions. The output is usually a set of items that cluster at the first layer, because that is what the unqualified word comprehension most readily denotes in the training distribution. A more productive practice is to specify the cognitive function the question is to perform, in the vocabulary of the activity-based approach, and to specify it separately for each layer. A prompt that requests questions requiring students to infer a speaker’s attitude from contextual clues, or to compare how two sections of the text handle the same theme, returns material at a level of specificity that a general request does not reach.

Specification by function rather than by content also makes the output easier to evaluate, because the teacher has stated in advance what a satisfactory question would have to do. This converts quality control from an open-ended judgement into a comparison between the question and the criterion that prompted it. Two further specifications are worth adding to function-based prompts. The first states the evidence base, by requiring that each question be answerable from a named part of the text. The second states the target of the question, by indicating the linguistic level at which students should be able to respond, since a question that requires production beyond the class’s current range is not a harder question but an unusable one.

The second strategy concerns the status assigned to what the model produces. If the output is treated as a finished set of questions to be delivered, the teacher’s role reduces to transmission and the lesson inherits the model’s defaults, including its preference for the obvious. If the output is treated as a pool of candidates, the teacher retains the decisions that matter. In practice this means generating more questions than can be used, then selecting. A useful working ratio, when preparing for a single reading-and-thinking text, is to generate three or four times as many candidates as the lesson can accommodate, and to select from them on the criteria developed in the following strategies.

Every question admitted to the chain must be anchored in textual evidence. At the first two layers, anchoring means that the answer can be reached by pointing to specific words in the text. At the third layer, where questions invite students to evaluate and extend beyond it, anchoring means something weaker but no less checkable: the text must supply the warrant for the response, in the form of the sentences, distinctions or constraints that a defensible position must answer to, while the position itself need not be fully determined by the text. This is the single most important filter, and it is also the one that GenAI most often requires, because models produce questions with the surface form of textual questions while their basis may be general knowledge, conventional expectation or invention [3]. The operational test is straightforward and can be applied in seconds: for each candidate question, name the sentence or clause that contains the evidence, and check that the answer follows from it rather than from what students are likely to know already.

Applying the test has a second benefit. It reveals the difference between questions that are about the text and questions that merely use the text as a starting point. Both have their place, and third-layer questions will often be of the latter kind, but a chain in which the earlier links are not anchored will drift, because there is no shared evidence for later questions to build on. It is worth noting that the anchoring test cannot be delegated to the model itself. A model asked whether its own question is supported by the text will usually affirm that it is; the check depends on a reader who has read the text carefully, which is precisely the professional act that the teacher supplies.

A chain exists when the answer to one question supplies something needed to answer the next. Making this relation explicit, rather than leaving it implicit, is both a design discipline and a teaching technique. As a design discipline it means writing the chain in a form that records, for each question, what it establishes and which later questions rely on it. Candidates that establish nothing and support nothing can then be removed immediately, which resolves the problem of abundance identified in Section 3.1. As a teaching technique it means that the dependency may be verbalized in class, so that students can see the structure of the enquiry they are being led through and can eventually reproduce it themselves.

Chains are most readily built along three axes. The evidence axis moves from what the text states, to what it implies, to what it leaves open. The stance axis moves from what the author thinks, to how the author signals that stance, to whether the stance is warranted. The transfer axis moves from the situation in the text, to an analogous situation the students know, to a position the students construct. Choosing an axis before generating questions is more productive than generating and then searching for an order, because the axis determines which candidate questions are compatible with one another.

Beyond generating questions, a model can be asked what students are likely to answer. Requests of the form what would a class of this level probably say to this question, and what misunderstanding would lead them to say it, return material that is directly useful in the lesson: the teacher enters the room knowing not only what to ask but what is likely to come back. Anticipated responses also refine the questions themselves, since a question that proves, on inspection of predicted answers, to have a single obvious answer is better replaced before the lesson than after it.

The most valuable outputs of this kind are plausible misreadings, particularly those traceable to a linguistic feature of the text such as a misleading cohesive chain, an ambiguous pronoun, or a metaphor whose habitual reading differs from its reading here. These are the points at which a lesson earns its time, and they are difficult for a teacher to enumerate unaided, because familiarity with a text erases the memory of what is difficult in it. GenAI has no prior familiarity with the text and is therefore free of this particular blind spot, but the claim should be read narrowly: what a model returns when asked to anticipate responses are hypotheses for teacher review, not observations of learners. Models may reproduce stereotyped or weakly supported assumptions about what students at a given level can do, and a predicted misreading that does not survive the teacher’s own reading of the text should be discarded. The model widens the teacher’s anticipation; it does not certify it.

The culminating question of a chain determines what the lesson has been for. Because it fixes the interpretive frame within which everything earlier is retrospectively read, it is the least delegable decision in the whole design. A model’s candidate interpretations are typically competent recombinations of established readings of a text and are unlikely to reflect what this class, with its particular history and its particular difficulties, most needs to be asked. The recommendation is therefore categorical: the teacher formulates the final question, and it is formulated after the rest of the chain is complete, so that it can gather up what the chain has established.

Table 2.The six strategies, their location in the design cycle, and the work each assigns to GenAI and to the teacher.

Reserving this question also preserves a feature of reading instruction that is easily lost in AI-assisted planning. What a text means is contested ground, and a lesson’s most valuable moments often arise when the teacher holds a reading that students must argue against rather than merely receive. A culminating question that has been formulated by the teacher and cannot be checked against a model’s answer keeps that ground in play. Table 2 summarizes the six strategies and the division of labour they assume.

The passage is a narration organized around a single theatre visit, and it is structurally simpler than it is conceptually rich. Before the performance the writer, a student who has read the play many times and seen several productions, is full of confidence. During it, one expectation after another is overturned: the orchestra opens with traditional Chinese instruments, the singing is in Chinese while music, movement and mime carry the meaning, the stage is reduced to a whip that represents a horse and a screen that represents a study, and the prince performs a backflip. After it, the writer concludes that a show combining music, singing, drama, poetry and costume with such effect proves that great drama crosses cultures.

The design problem mirrors this structure. A lesson that stays with the sequence of events will be fluent and unremarkable, because students can follow a vivid account easily. The real demands lie elsewhere: in the writer’s arc of feeling, which runs from confidence through curiosity to astonishment and is signalled by evaluative language rather than stated as a thesis; and in the title’s metaphor, which frames the meeting of a Western classic and a Chinese art form as an encounter rather than a contest and never argues the point directly. The chain below is built along this stance axis.

The first layer is deliberately brief: three questions, each establishing ground that a later link uses. The first asks how the account is organized in time, which establishes the before-during-after frame. The second asks which features of the performance the writer describes, which establishes the inventory of details, from the jinghu and the singing voices to the stage properties and the movements. The third asks how the writer’s feelings change across the visit, which establishes the affective line that the second layer will explain.

The design decision worth noting here is what was excluded. A full pool of first-layer candidates would include questions about the plot of Hamlet, the history of Peking Opera, and the meaning of individual props, each answerable by inspection or by general knowledge. None of these establishes anything a later question depends on, and under Strategy 4 they are removed. Their removal is what creates the time for the higher layers, and it is the clearest illustration of the point made in Section 3.1 that the teacher’s work at the first layer is subtractive.

The second layer asks students to reason across the ground the first layer has established. A representative question asks why the writer, who begins in full confidence, ends the evening on the edge of the seat, which requires students to connect the affective line with the details that produced it. A second asks what the writer’s comparison of the two art forms implies about their differences, an inference grounded in the account of the singing, the movement and the stage. A third asks students to identify the expressions with which the writer signals his astonishment, and a fourth asks what is meant by saying that the performance transformed a small stage into the whole universe, which moves from an unusual metaphor to the conception of theatrical space behind it.

The second and fourth items are the kind that GenAI produces readily when the prompt specifies the function, and they are also the kind that must be checked against evidence under Strategy 3. In this text the writer’s stance is carried by a small set of evaluative expressions whose placement is not accidental; a candidate question that asks for the writer’s attitude without directing students to those expressions is not anchored, whereas a question that asks students to locate them is. The difference between the two is the whole of the strategy.

The third layer opens the text outward. A first question asks whether the account supports the claim that a traditional art form must be renewed before it can hold a contemporary audience, and requires students to decide what in the text bears on the question. A second asks students to weigh what is gained and what is lost when a classic is transposed into another art form, comparing the Peking Opera Hamlet with versions of the play they know. A third, formulated by the teacher rather than generated, asks why the title stages the encounter as when Hamlet meets Peking Opera rather than the reverse, what that direction assumes about the two traditions, and whether the class finds the assumption convincing.

The last question is the culminating question of the chain, and it is deliberately not one that a model would be likely to propose. It does not ask for a verdict on the performance, which is the conventional move at this layer, but for an examination of the direction in which the text points the encounter, an examination made possible only by the distinctions that the earlier layers have established. Students who reach it have travelled from an inventory of performance details, through the writer’s changing stance towards them, to a claim about how the text conceives of cultural exchange, and each step of that journey was prepared by the step before it.

Read end to end, the chain has ten questions, of which three operate at the first layer, four at the second and three at the third. It is worth observing how this distribution came about. Under Strategy 2, a pool of roughly thirty candidates was generated and roughly a third of them were retained, with the largest attrition at the first layer. Under Strategy 3, several otherwise attractive first- and second-layer items were discarded for lack of a checkable evidence base, particularly questions about the plot of Hamlet and the history of Peking Opera that the text touches but does not develop. Under Strategy 4, the survivors were arranged along the stance axis and the dependencies were recorded. Under Strategy 6, the final question was written by the teacher after the rest of the chain was fixed.