A textbook exercise can look correct, read pedagogical, and teach nothing beyond the Claude documentation. It will produce the habit of using LLMs like a phrasebook: correctly, generically, without understanding what the words mean in context.
The fluency trap is a repeating shape. In Chapter 1, we saw how Claude could produce fluent copy that sounded like design writing without understanding the brief. In Chapter 9, we saw how fluent-looking layouts could be generated without the designer's judgment—without understanding what mattered visually. Here the same pattern emerges at pedagogy. An exercise can look pedagogically sound, ask the student to do something, produce a deliverable, and still teach nothing that matters. The danger is that the student completes the exercise, feels competent, and never realizes their learning is still generic, still divorced from the practice they actually do in their field. They have learned to use the tool, not to integrate it into their domain.
This is not a new problem. Eric Mazur's research on concept inventories showed that students who pass traditional exams routinely fail at the application layer when the application is novel. Generic LLM exercises produce the same failure curve.
If an exercise could appear unchanged in a textbook on cooking, accounting, screenwriting, or veterinary medicine, it has not distinguished your domain. An accountant has no client brief in the sense that a designer does. A veterinarian has no design project. An educator has no brief at all. The exercise that works for all fields works for none. This is the stakes: the student learns to use the tool, but not to use it where it matters—in the specific, irreducibly human layer of their own practice. They leave the textbook with a tourist's phrasebook, not a practitioner's understanding. And because the exercise was structured correctly, because it looked pedagogical, neither the student nor the instructor recognizes the gap.
Exercise A is generic—it works in cooking, accounting, screenwriting, veterinary medicine. Exercise B is irreducibly specific: it assumes a real brief, a real client, a real history only the designer has built.
These two exercises from the ai-for-designers draft illustrate the difference. Exercise A asks the student to understand structure—how to explain it to Claude, how to evaluate Claude's execution. It would work for any field. Exercise B cannot transplant. It is built entirely around the specific, irreducibly human layer: the designer's history with a client, their calibration of that client's voice, their accumulated reads of feedback across projects. The exercise asks Claude to serve as a second reader, and the data comes from disagreement. The designer brings something only they have—client context, embodied expertise, relational knowledge. Claude provides a fresh read. The learning happens in the friction between them. An accountant has no brief that works this way. An educator has no brief at all. The exercise is only useful for this domain, this reader, this career stage.
This is what distinguishes domain expertise from generic tool use. The model cannot learn it from text. Only the practitioner has it.
The irreducibly human layer is the domain expertise that the model cannot infer from training data because that data is either private—your client relationship, your project history—or embodied—your calibrated sense of what matters in your field—or local—the specific friction points in your practice. When a designer asks Claude to read a brief as a senior creative director would, Claude is being used not as an expert on that client but as a second reader. The friction between Claude's read and the designer's read is the data. Claude misreads because the model has not worked with this client before. Claude sees fresh patterns because it is not burdened by the history. Both the misreading and the fresh seeing teach something. But an exercise that does not involve a real brief, a real client relationship, a real accumulated history cannot access this layer. It remains generic, which is to say: it remains in the layer where the model is already competent.
The different field is at the field level, not the sub-specialty level. A graphic design exercise that transplants to product design might still pass—they share workflow structures. A graphic design exercise that transplants to accounting fails. The boundary is where the irreducibly human layer lives.
The field-boundary test determines whether an exercise has accounted for domain expertise. This is precise: the standard is not individual-distinguished, not sub-specialty-distinguished, but field-distinguished. A graphic designer's exercise could theoretically work for product design because both fields involve visual iteration with stakeholders, both involve critiquing mockups, both involve client feedback. But an exercise that works for graphic design and accounting does not—cannot—work. The irreducibly human layer in accounting is entirely different. The model of client engagement is different. The feedback loop is different. The embodied knowledge required is different. When an exercise passes the field boundary test, it means the designer has recognized that their domain is not the same as every other domain. They have built the exercise so it could not work for someone in a different field, no matter how skilled that person is, because it requires domain knowledge and practice that only the designer's field develops.
The AI+1 standard is engaged pedagogy translated into an exercise-design rule. Paulo Freire's critique in Pedagogy of the Oppressed and bell hooks's extension in Teaching to Transgress both argue that learning without context is not learning—it is banking.
Paulo Freire's critique of the banking model—the teacher deposits content, the student stores and withdraws—is that it treats the learner as a container for transferable knowledge rather than as a domain-situated practitioner. A generic LLM exercise is a banking-model exercise. It deposits generic prompting technique and assumes the student will be able to withdraw it later in their domain. But the withdrawal never happens the way it should. The student has learned the technique in a generic context, without the friction, without the stakes of their own practice. Bell hooks extended Freire in Teaching to Transgress, arguing that pedagogy must be engaged—must take the learner's specific context seriously as the precondition for competence, not as decoration around it. The AI+1 standard is engaged pedagogy operationalized into a rule: an exercise has to test domain integration, not prompt syntax.
The concept-inventory tradition in physics education has repeatedly demonstrated that students can pass summative exams without developing competence at transfer. The exercises that work are the ones that embed the learning in the authentic context where the knowledge has to be deployed.
Eric Mazur's research at Harvard on concept inventories and peer instruction found a critical pattern: students who pass traditional physics exams routinely fail to apply that knowledge when the problem is presented in a novel context. They have learned the syntax of the solution, not the concept underneath it. They can follow the steps in the problem set but cannot recognize when those steps apply in a new situation. Generic LLM exercises produce the exact same failure pattern. A student who completes Exercise A—explaining structure to Claude, evaluating Claude's output—can do that task. But when they sit down with their own project, their own unfamiliar domain, their own client brief, they have no template for how Claude fits into that work. They have not practiced integration. They have practiced generic prompting. The only way to practice domain integration is to do it in a domain-specific context, with stakes that matter in that domain.
An exercise that passes all three is doing AI+1 work. An exercise that fails any one is, on at least one axis, generic.
The three questions form a coherent test. The first asks whether the exercise is domain-specific. The second asks whether the exercise accounts for what the practitioner uniquely brings—embodied knowledge, client relationships, project history. The third asks whether the exercise treats the LLM as a tool that the practitioner uses to make a judgment, or whether it treats the practitioner as an evaluator of what the LLM already produced. If a student is just reading Claude's output and judging it good or bad, they are not integrating Claude into their practice. They are doing literary criticism of Claude, not design work with Claude. The deliverable has to be something the reader could not produce without engaging with Claude, and something that requires the reader's domain expertise to create. A designer asking Claude to read a brief and then using Claude's insights alongside their own accumulated client knowledge to make revision decisions—that is a judgment the designer produces. Reading Claude's summary of the brief is not.
When the enrichment generator drifts toward hyper-specificity, edit back. The exercise should be general enough for any practicing designer, specific enough that it could not work unchanged in accounting.
The field boundary is the right level of specificity. Too far in one direction—the generic direction—and you produce banking-model exercises that teach nothing. Too far in the other direction—the hyper-specific direction—and you produce exercises that only fit one reader, one client type, one specialization within a field. The bound is set at the field level because that is where coherent bodies of practice exist. Graphic designers, across all their specializations, share certain methods: they iterate on visual mockups, they critique with stakeholders, they manage feedback and revision. The exercise should be general enough to work for someone specializing in packaging as well as someone specializing in branding, in web design as well as in editorial. But it should not be general enough to work for an accountant. The irreducibly human layer—the client relationship, the accumulated reads of what works in the client's industry—exists at the field level, not the sub-specialty level and not the individual level.
The three fluency traps—verbal, visual, pedagogical—are instances of the same underlying risk: that a tool becomes so fluent it becomes invisible, and that invisibility allows incompetence to hide as competence. Fighting this requires building exercises from the ground up: from what the student actually does, from what they uniquely bring to the work, from the judgments they have to make in their domain.
The synthesis of this chapter is straightforward: pedagogy cannot be separated from domain. A textbook on design pedagogy for AI tools must be a design textbook, not a prompting textbook. The field of graphic design has a coherent body of practice: how to work with clients, how to iterate on visual concepts, how to receive and interpret feedback, what makes a communication effective or ineffective in a specific context. An LLM is a tool that can be integrated into that practice in certain ways—as a second reader of a brief, as a generator of visual directions to critique, as a way to explore variations faster. But the integration happens in the domain, not in a generic space. When a textbook teaches the generic technique and hopes students will integrate it later, it is practicing the banking model. When a textbook builds the exercise from the domain—from the real client brief, the real accumulated knowledge, the real stakes—it creates the conditions for genuine learning. The exercise becomes a practice of the craft, not a lesson in a tool.
The model contributes fluency, breadth, and fresh perspective. The reader contributes judgment, context, and domain expertise that no training data could supply. When these meet in an exercise built from the irreducibly human layer of the field—the client relationship, the embodied knowledge, the practice-specific stakes—then the student learns not to use a tool, but to integrate it. That integration is everything.
The thesis connects all three levels of the fluency trap. Verbal fluency occurs when words sound right without understanding context. Visual fluency occurs when layouts look right without understanding hierarchy or intention. Pedagogical fluency occurs when exercises look right without accounting for domain integration. The AI+1 standard breaks this pattern by making the irreducibly human layer—what only the practitioner brings—an explicit requirement. An exercise that could work for someone in another field has not met this requirement. An exercise that does not require the reader to bring something only they have has not met this requirement. An exercise that delivers just LLM output rather than reader judgment has not met this requirement. When all three are addressed—when the exercise is field-specific, when it requires the irreducibly human layer, when the deliverable is the reader's judgment—then the relationship between the model and the human inverts. The human is not evaluating what the model produces. The human is integrating what the model produces with what only they can know. That integration is the learning. That is what distinguishes competence from fluency.
AI 1 · Chapter 10 · Enrichment for AI