1. THE PROMPT
I recently saw the following prompt recommended for teachers:
At first sight, this prompt looks quite reasonable. It gives the subject, the type of students, the expected knowledge or skills, and the course plan. It also asks for a clear structure.
But it is a very bad prompt.
The problem is that it asks the AI to perform serious knowledge work without telling it how this work must be performed. This prompt tells the AI what product we want, but it does not give the AI the method by which this product must be created and therefore it invites LLM to fill this gap with LLM’s own preferences – without even inquiring what those preferences are.
2. THE MISSING METHOD
When we ask an AI to “suggest learning goals,” we leave several important questions unanswered.
What is a learning goal? How is it different from a learning outcome, an objective or a competence? Which educational theory should be used? How should the goals be connected with assessment? What makes a goal good, complete or measurable?
The AI must answer these questions somehow. If we do not provide the answers, it will take them from what it learned during pre-training.
This means that the model itself will choose the method. It may use Bloom’s taxonomy, or not. It may use the competence model that is the most popular on X but not one that is actually sound and trusted. It can combine several traditions in a way that no one predicted. It may follow common patterns from syllabi of some obscure university with artificially raised search rank.
The result will sound professional, it will include active verbs and a clean hierarchy. Yet this doesn’t prove that the method was correct.
Before a measurement device is used for an important measurement, it must be checked and calibrated. We do not trust its reading only because a number appears on its screen. The same principle should apply to AI: before we use it for knowledge work, we must calibrate its way of working. Nothing important should remain at the discretion of knowledge hidden inside the model. The human who orders the work should understand and approve the method.
For this reason, the proper starting point is an accepted book, instruction, standard or professional guideline about designing learning goals. The human should read this document first and decide that its method is suitable. Then the same document should be given to the AI. The AI should work from this document, not from an unknown mixture of material inside its pre-training.
3. THE COST OF A HIDDEN ERROR
Learning goals are not a decorative part of a course description. They influence what will be taught, what will be omitted, how students will be assessed and what the institution will call successful learning. If the method for designing the goals is wrong, the error may spread through the whole course.
A grammar mistake is easy to notice. A hidden methodological mistake is more dangerous.
The original prompt gives the AI two kinds of authority. It allows the AI to choose the method and then to produce the answer by that method. The teacher receives only the final result and does not see the theoretical choices that produced it. In such a situation, proper verification is almost impossible. We can say that the result “looks good,” but this is not a real test. We cannot compare it with an approved method because no method was selected in advance.
The output may still be useful. But if it is useful, this may be partly a matter of luck. Luck is not a proper basis for curriculum design.
4. NAVIGATIONAL THINKING
Professor Ilya Levin’s concept of Navigational Thinking explains why this prompt fails. It explains it more precisely than a general complaint about missing detail.
Levin starts from the way a language model works. The model does not store facts and then retrieve them. It places every concept as a point in a space of very many dimensions. It produces an answer by moving along a path through that space. The path is the work. The answer is the place where the path ends.
This changes the question we should ask about meaning. A symbolic system asks where a concept is located. This kind of space asks in which direction a concept points. Levin states the shift in two words: from “where?” to “whereto?”
Navigational Thinking is his name for the thinking suited to such a space. One feature of it matters most for our case. The problem space is not known before the work starts. It is constituted during the work itself.
The teacher’s prompt gives the model a destination and nothing else. It says, in effect: “Go and produce learning goals.”
A destination alone is not enough. The prompt provides no map, no calibrated compass and no marks by which the route can be checked. The model will still move. It holds enough material from pre-training to choose a direction. But the route is selected inside the model, and we do not see it.
The geometry of the space makes this worse than it first appears. The number of possible routes is enormous, and many of them end at answers that read well. A polished result therefore tells us very little about the method that produced it. We cannot recover the method from the text.
The human should choose the map before the work starts. The human should also fix in advance the marks that show whether the AI is moving in the right direction.
Levin also assigns the human a definite task. The model supplies the geometry. The human supplies the direction. This direction comes from professional experience that is difficult to put into words. An experienced teacher often knows that a set of learning goals is wrong before she can explain why. This knowledge is valuable. It is wasted if the work is arranged so that it can only be applied at the end.
An approved methodology solves this problem. It converts private judgement into a public standard. The human can then point to a rule and show that a goal breaks it, instead of saying only that the goal looks weak.
Levin adds one more point about the order of work. He treats navigation and formal method as two phases in sequence, not as competitors. Navigation comes first, while the problem still has no stable form, and it establishes that form. Formal method comes second and operates inside it. The teacher’s prompt reverses this order. It asks for a finished product while the method is still undecided.
Understood in this way, the concept requires an exchange instead of a single instruction. The system proposes a direction, the human examines it. Disagreements become visible, and the direction is corrected. Work continues only when the route is clear enough to defend. One large prompt cannot carry such an exchange.
5. AGENTIC SOFTWARE ENGINEERING
In my work on Agentic Software Engineering, I describe a discipline based on four connected activities: (1) frame, (2) specify, (3) build and (4) verify. The same discipline shall be applied outside software engineering.
First, we frame the work. For curriculum design, this means describing the course, the students, the institutional setting, the available time and the real educational purpose.
Second, we specify the work. Here we give the approved methodology for writing learning goals. We define the important terms. We decide what form the goals must have and what rules they must follow.
Only after this do we build. At this stage, the AI may create a first version of the learning goals.
Finally, we verify the result. Each goal should be checked against the course material and against the selected methodology. We should be able to explain why the goal exists and which rule supports its form.
The sequence is important. If we ask the AI to build before we specify, we are asking it to invent the missing specification for us. The reference book or guideline is therefore not additional reading, it is part of the working specification, it tells the AI which method it is authorised to use.
The human remains responsible for accepting the method and the final result. The AI helps to perform the work, but it does not silently define what correct work means.
6. A CONTROLLED WORKFLOW
The safest workflow begins with an approved source: a book, a guideline, a methodology document.
First, the human selects the file, reads it and accepts its approach.
Then the course plan and this methodological document are given to the AI (attached to the prompt). At this point, the AI should not design any learning goals. Its first task should be to explain the method it found in the document. It should list the definitions, rules, steps and quality criteria. Whenever possible, it should give page references.
The human then reviews this explanation. If the AI misunderstood the document, the misunderstanding must be corrected. If the document leaves an important question open, the human must make a decision.
After this, the human and the AI agree on the required output. They decide how many goals are needed, how they will be grouped and how each goal will be checked. Only now should the AI produce the learning goals.
The finished goals should then pass a verification stage.
For example, each goal can be connected with:
a part of the course plan;
a rule from the approved methodology;
an expected student capability;
a possible form of assessment.
Sometimes the human does not have a suitable guideline. In this case, the human should first interview the AI.
The AI should be asked which framework it proposes, where that framework comes from, how it defines a learning goal, which steps it plans to follow, which alternatives exist and what weaknesses its proposed approach may have.
The human should discuss these answers with the AI. Unclear points should be clarified. Unacceptable assumptions should be changed. The discussion should continue until the human understands and accepts the method. This accepted method should then be written down as a temporary specification. Verification criteria must also be agreed upon.
Only after that may the AI start producing learning goals.
The lesson is simple: the correct unit of work is not a “ready-to-use prompt.” It is a controlled workflow. Before we ask AI to create an important result, we must know which method it will use. We must approve that method, and we must decide how the result will be checked.
We calibrate the instrument before we trust the measurement. With AI, calibration means an approved source, a visible procedure and a result that can be verified.