During the summer we worked on a book chapter about using AI to improve student reading skills, we had just finished working with Chain of Density (CoD) to improve student reading (see our post on that here), and right in the middle of this Eugenia had a brainstorm. She called me: “We have used AI to augment writing, to tutor our students, and to edit writing—but we haven’t used it to actually improve writing. Is it possible to do that?”
It was a good point. We hadn’t really used AI to its full potential, and we had a long way to go in this AI journey. The idea of using AI to actually improve student writing was a great idea and a wonderful challenge (and who doesn’t love a challenge?) Because we had just finished working with CoD, I started to play with some ideas from the “chain of thought” (CoT) prompting school where the CoD prompt had come from. I wondered if we could use CoT to do exactly what Eugenia was talking about.
What if, instead of our students prompting the GenAI to enhance their paragraphs, the GenAI was used to prompt the students to enhance their paragraph? Then, the writing would be done by the students, with the AI pushing them to improve. Using CoT prompting, I made a far-too-constrained prompt that focused on descriptive writing. It worked great, I have to say, but as we continued to brainstorm and talk about possibilities, we also talked of a prompt that could improve research paper paragraphs—especially those paragraphs that included a PEEL structure and needed work integrating sources. I started working on a similarly constrained prompt that would do research paragraphs.
Then it struck me. I really didn’t need to make the prompt that presecriptive? I tried loosening the language in the prompt, making it more general, and then I found that it worked for both types of paragraphs. It was very exciting. We were going back and forth about how we could integrate this into our already existing course, but we were too far into the semester to really pilot it.
So, we piloted it last week in our English 1101 to help our students improve their descriptive writing. We also put it into our English 1102 course to help the students improve the integration of sources into their research papers.
This is how it works: The student starts by writing a short paragraph (around 100 words). Then they paste the prompt into ChatGPT, You.com, or Claude (for some reason it does not work in Gemini—we have no idea why). The GenAI analyzes what they have written, then asks the student to add more detail of a specific type. For example, here’s a transcript of the prompt (in italics) and an example of how it works with a paragraph I wrote. My paragraph is meant to “show” the sentence, “He was tall”:
MK: Please provide your original paragraph below, and I'll wait for you to enter it.I will analyze your paragraph and suggest one specific area for improvement. Based on my suggestion, please add a new detail or description to enhance your paragraph. I will incorporate your new detail into the paragraph and display the updated version. We will repeat steps 2-3 four more times, focusing on a different aspect each round. After five rounds of enhancements, I will present the final version of your improved paragraph.
CLAUDE: Please provide your original paragraph below:
MK: Tom stooped a bit to get into the front door of my home, and as soon as he entered, the room seemed full. It wasn't that he was broad, it was just that everything else seemed so tiny. I wanted to offer him a place to sit down, but all the chairs seemed inadequate. I chose, instead, to grab my purse. "Thank you for coming to get me," I said a bit too forcefully. "Let's get out of here."
CLAUDE: Thank you for providing your paragraph. I'll analyze it and suggest an area for improvement.
The paragraph does a good job of conveying Tom's size and the narrator's discomfort, but we could enhance the sensory details to make the scene more vivid. Let's focus on adding some visual details about Tom's appearance.
Could you provide a new detail or description about Tom's physical appearance that would help the reader better visualize him?
MK His nordic features, blonde hair and chisled face, rose above a massive chest and broad shoulders. His blue eyes blazed.
CLAUDE: Thank you for that vivid description. I'll incorporate this new detail into the paragraph. Here's the enhanced version:
Tom stooped a bit to get into the front door of my home, and as soon as he entered, the room seemed full. His nordic features, blonde hair and chiseled face, rose above a massive chest and broad shoulders. His blue eyes blazed. It wasn't that he was broad, it was just that everything else seemed so tiny. I wanted to offer him a place to sit down, but all the chairs seemed inadequate. I chose, instead, to grab my purse. "Thank you for coming to get me," I said a bit too forcefully. "Let's get out of here."
Now, let's focus on another aspect of the paragraph. The setting could use more detail to create a stronger atmosphere. Can you provide a new detail or description about the narrator's home or the room they're in? . . .
The enhancement goes on for five steps. Each time, the GenAI asks for a different type of enhancement to make the paragraph better. We couldn’t wait to pilot the exercise. We tried it with all types of paragraphs, not just descriptive paragraphs, and it worked beautifully.
The Pilot Results
We were pretty stoked that this prompt seemed to work so well, so we piloted it in our English 1101 classes last week. In the assignment, students were to provide the original paragraph, the final enhanced paragraph, a link to their chat session, and a short reflection on how it went.
Most of our students got a lot out of this exercise, and really learned how to add things to their writing that would improve their descriptions. However, this was not always the case.
With a few students in each class, we were reminded that piloting something you are enthusiastic about with a class full of freshman students is like a lot like being plunged into a very cold bath.
We had not anticipated the errors our students would make, so our directions were not comprehensive enough. In addition, we also had problems with students’ prompt adherance:
Student Error #1: The Conscientious Students
Our very conscientious students wrote the additional information when prompted, then proceeded to integrate that additional information into a paragraph, then pasted the whole paragraph in. This extra work and effort was unneeded, and it kept the prompt from working properly. They should have simply followed the GenAI’s directions and the GenAI would have integrated the information into a paragraph for them. We did not clearly explain that students didn’t need to keep pasting in the whole paragraph because we didn’t anticipate that it would be a problem.
Student Error #2: The Lazy Students
These students didn’t want to do the work required to improve their writing. So, they opened another tab with a different GenAI and asked the new GenAI to produce the additional information, then pasted that in answer to the prompt. This resulted in a patchwork paragraph that had no original student voice, and ended up being narrative rather than descriptive. We uncovered this by looking at the reflection paragraphs that clearly differed from those of other students, and the strange chat history they shared. We asked these students to repeat the exercise with their own writing.
Student Error #3: The Prompters
These students were either trying to turn the exercise back around to having GenAI write the paragraph, or they were just confused about how to make this work (We couldn’t really decide). These students added prompts to the chat, (one student even asked GenAI to write the sentence for him), or they answered the GenAI’s prompts by telling the GenAI to do something else. This broke the prompt, and the entire exercise was moot. We discovered these students, again, by looking at their atypical reflections and examining their chat histories. We gave them the benefit of the doubt (well, except for that student who told the GenAI to write the sentences for him), and we asked these students to try again.
Outcomes
So, the next time we do this, we will remind our students that they must sign into the GenAI they are using so that they can link the chat (a lot of our students forgot to do this). We will also adjust our directions to explain to the students that they are not supposed to post anything but the information they have been prompted to include, that they do not have to integrate their additional information into a paragraph (the GenAI will do that), and that they should not add any prompts.
As we were grading the work, we discussed the purpose of the assignment. Were we teaching writing, or were we teaching how to use GenAI? We disagreed. We were teaching writing, I argued, because we had them write the enhancements themselves, and they learned what they were missing to make the paragraph as vivid as possible.
Eugenia disagreed. She said that we were teaching the students how to use AI properly, but that was an important lesson in itself.
However, we both agreed that it was a good exercise, no matter what we thought about its purpose. Our students learned both things, really—how to use AI better, and how to write better—and, of course, we learned the significance of piloting something carefully because you never know how it will go.
We will be doing this again, and again, and again. It was fun, and it was helpful. Of course, it doesn’t mean that we won’t get something entirely different in the next iteration—but isn’t that why we all love to teach? Our students definitely keep us on our toes!