‘The more I improved my instructions, the better they worked as AI prompts’

Clear assignment instructions can be understood as an act of care. They tell students the goal, the context, the required steps, the learning outcomes and the criteria for success. They reduce the advantage enjoyed by learners who already know how to decode the hidden expectations of academic work. 

I believed in this picture, and I still do.

Then, this fall, I uploaded one of my own carefully revised assignments to a chatbot and typed three words: “Complete this assignment.” What came back in a minute scored 48.5 out of 50 against my own rubric. What hit me hard was the realisation that the more I improved my instructions, the better they worked as artificial intelligence prompts.

The assignment and the prompt

An assignment sheet and its rubric are a natural-language specification. They tell students what output to produce, which materials to use and which qualities will be rewarded.

An AI prompt does much the same thing. 

• Don’t ban GenAI: teach students to question it
• If a chatbot can pass your assignment, you are grading the wrong thing
• AI’s fluent answers hide bias. Here’s how to teach students to see it

Suppose I ask students to analyse two readings, apply a framework introduced in class, address a plausible objection and support their conclusion with textual evidence. My rubric explains what excellent analysis and strong evidence look like. For a student doing the work, those instructions make the intellectual path easier to see. For a large language model (LLM), they make the desired output easier to generate. (This is no surprise since producing fluent, well-formed text is what these tools are supposed to do the best.) The student may not even need to write a sophisticated prompt. “Complete this assignment” is enough once the assignment instructions, rubric, readings, lecture recordings, slides and notes have been uploaded. The carefully designed materials become the carefully written prompt.

This is what I have come to call the clarity paradox.

Clarity has two audiences now

Before generative AI, assignment instructions had one primary audience: students. Clearer instructions reduced uncertainty and anxiety; clearer rubrics made grading fairer and more transparent. Those instructions now have a second audience: a model capable of producing fluent, rubric-responsive output.

This does not mean that generating output is equivalent to learning. The point is more limited. For the learner, specificity lowers the cost of understanding what is expected. For the AI user, the same specificity lowers the cost of producing something that looks as though those expectations have been met.

The complication is easy to miss because clarity and AI-resilient design are often recommended together. A guide from Rochester Institute of Technology’s Center for Teaching and Learning, for example, advises transparency when designing assignments with AI in mind. Under Articulate Criteria for Success, it states that “clear criteria redirect students toward deeper thinking and away from AI shortcuts”. The same guide asks instructors a diagnostic question: “Have I redesigned the task so that the learning cannot be outsourced?” These sentences appear in the same resource, a few clicks apart.

Both goals are worthy. The relationship between them, however, is not straightforward. Clarity can reduce ambiguity while increasing technical executability.

A better product is not always better evidence

The question is not whether an AI-generated paper is good. It is what the paper allows educators to infer about students and their learning.

When I scored that fully AI-generated submission, I graded it the way I grade everything: reading for the qualities my rubric names. It earned nearly 97 per cent. Then I opened the links for references. Of the five sources it cited, one was correct. Two had invented titles. Three had links that did not work. One was a real government dashboard cited for data it did not contain, beneath a fluent paragraph interpreting numbers no one had read. Scored again with every source checked, the same submission’s mark fell to 79 per cent.

Every one of the nine points that separated the two readings sat in the criteria that depended on whether the text corresponded to something out there. That is the distinction worth holding on to. Criteria that grade the properties of a text – structure, coherence, completeness, tone – can be satisfied without warrant. Criteria that grade the relation between a text and the world cannot.

It also clarifies what we can reasonably demand. I am not trying to determine who or what wrote a submission; my course permits AI use under stated conditions, and most student work now falls somewhere between fully authored and fully generated. What a student owes for any claim they submit is warrant: the capacity to say why it is asserted, what supports it and how it might fail. That obligation is indifferent to how the text was produced. A claim generated by a model and then verified is warranted. This is often promoted as a good, responsible AI use. A claim authored by a student but based on a misremembered source is not.

Fabricated citations are not primarily a failure of authorship. They are a failure of warrant. Unlike authorship, warrant is something we can check in this AI era.

Questioning clarity is not an argument for vagueness

One tempting response would be to make assignments less explicit: to withhold criteria, remove examples and introduce ambiguity, so AI has less to work with. That would be a mistake. Confusing students is not a strategy for protecting learning, and it would disadvantage novices, multilingual learners, students with disabilities and anyone unfamiliar with academic norms. 

The better response is to separate two questions that are often put together:

• How can I make this assignment more accessible?
• How can I obtain credible evidence that student learning occurred?

Clarity addresses only the first, while the second is concerned primarily with AI resilience.

Sometimes one design feature serves both. Harvard’s Derek Bok Center for Teaching and Learning recommends pairing take-home work with in-person checkpoints. Methods such as introducing a brief in-class oral assessment and short live demonstration can be transparent and still require students to demonstrate their understanding. 

Adding more specifications to a take-home assignment is different. Frequently mentioned methods such as process logs and personal reflections may support learning, but they can also be generated without learning. Documented process is still a product.

So, the test is not whether an activity sounds reflective, authentic or specific to a course. It is whether a student can satisfy its visible requirements without doing the cognitive work the assignment is meant to elicit.

From instructional clarity to evidentiary clarity

Not every assignment needs an oral examination, and not every course can sustain one-to-one conversations. Instructors already face real limits on time, class size and institutional support. I am struggling to see what methods I can adopt sustainably.

But naming the clarity paradox lets us choose more deliberately. For each significant assessment, we might ask three questions:

• What cognitive work do I want students to perform?
• Could AI produce the evidence my rubric names without students doing that work at all?
• What small change in the conditions of performance would strengthen the connection between the work and the learner?

The goal is not to make every assignment impervious to AI, which is not achievable nor desirable. The goal is to know what our evidence can support.

I will keep trying to make my assignments clear. Students deserve to know what they are learning and what the intended learning looks like. But I can no longer assume that a clear description of successful work will, by itself, lead students away from AI shortcuts. In the two years I’ve taught critical thinking to undergraduate students, I have been making my instructions better and, without intending to, making them easier to hand to a machine.

Clarity still matters. It has simply stopped answering the question I thought it answered. The one I have to ask now is narrower and harder: what makes this submission evidence of this student’s learning?

I do not have a general answer. But I have stopped mistaking a well-designed assignment for one.

Soohyun Ahn is an instructor in philosophy and course lead for critical thinking at the Southern Alberta Institute of Technology, Canada.

If you would like advice and insight from academics and university staff delivered direct to your inbox each week, sign up for the Campus newsletter.