Image created by Heidi K. Brown with ChatGPT’s DALL-E

Enhancing GenAI Writing Output Through “See One, Do One, Teach One”

Daily marketing messages flood our psyches lauding GenAI’s swiftness in generating written content: Ramp up efficiency! Produce faster results! Accelerate your workflow!

As writers, we might feel like we’re suddenly expected to generate all our written work at the speed of light. But obviously, velocity doesn’t necessarily equate to quality. Instead of succumbing to feeling dazzled, dizzied, or disoriented by GenAI’s lightning-fast output, let’s give ourselves permission to decelerate before accelerating our workflow. Let’s (at least temporarily) ease up on the throttle that’s materialized in our hands, start basic, gently introduce ourselves to GenAI, and design a methodical step-by-step process for using these tools to generate high-quality writing. Like any new relationship or situationship, it takes time to learn how to best communicate with a new partner or pal. By practicing GenAI communication skills, we’ll get better at understanding how our input drives the quality of GenAI’s output. One possible strategy: applying the medical profession’s teaching-learning model of “See One, Do One, Teach One” (SODOTO) to the GenAI writing process.

Dr. William Stewart Halsted, the first surgeon-in-chief and founder of the surgical residency program at Johns Hopkins, designed a three-part process of teaching and learning:

• See One: A learner observes how an experienced practitioner performs a task
• Do One: Under supervision and guidance, the learner performs the task themselves and receives (and processes) constructive feedback
• Teach One: After developing sufficient competence at the task, the learner teaches the incremental steps of the task to another learner

Using the SODOTO model, let’s experiment with priming GenAI tools to help us produce the high-quality written content we aspire to create.

See One: Examine Models of High-Quality Writing within a Genre

Whether we’re novice or veteran writers, or have little or lots of comfort with technology, let’s grant ourselves permission to adopt a beginner’s mindset toward GenAI+Writing. As GenAI newbies, let’s engage in an activity called “genre discovery” — the writer’s version of “seeing one.”

Linguist John Swales defines a “discourse community” as a group of people who use “writing genres” to communicate with one another. (For more on discourse communities, please see this great book chapter by Professor Dan Melzer.) Professor Katie Rose Guest Pryal encourages writers stepping into new discourse communities to engage in “genre discovery.” Professor Pryal and Professor Kerry Dirk advocate for collecting, studying, and dissecting samples of a genre document to identify common “conventions” — structural components, stylistic features, vocabulary, citation form, and more. Then, Professor Susan Tanner recommends we capture this acquired knowledge by creating “how-to manuals” for each new genre we discover.

Activating the “see one” step of SODOTO in the context of GenAI, consider the following genre discovery progression:

• For novices within a new discourse community, seek out exemplars of the genre of writing you plan to (eventually) use GenAI to create. For veterans, choose exemplars of the genre from your own writing portfolio.
• Next, discern what makes each exemplar “good” or “bad.” In other words, identify the writing criteria that qualifies an exemplar as “high-quality.” (Novices who are not sure what makes an exemplar “good” or “bad” can seek guidance from more experienced writers within the discourse community.)
• Now, select two or three “good” exemplars to use as genre models.
• From the exemplars, identify the structural building blocks, stylistic features, and other characteristics (like vocabulary, citation form, etc.) that all your “good” exemplars within the genre have in common.
• Finally, create a “rubric” of criteria for evaluating documents (and their building blocks or component parts) within the genre.

“See One” Example: A law student or junior attorney tasked with writing a legal memorandum could engage in the “see one” step of SODOTO by:

• researching and collecting sample legal memoranda
• deciding which samples are well-written, weeding out those that aren’t (asking for guidance by more experienced memo-drafters as needed)
• identifying the common building blocks or component parts of a well-written legal memorandum (e.g., Memorandum Header, Question Presented, Brief Answer, Statement of Facts, Issue Statement, Rule, Explanation-of-the-Rule, Application-of-the-Rule, Conclusion, clear headings, proper legal citation, etc.)
• creating a “rubric” of criteria for evaluating each component part of a memorandum

Do One: Practice Writing a Genre Document (or a Genre Building Block)

Next, let’s engage in SODOTO’s Step Two: “doing one.” The step of “doing one” — the cognitive process of writing a genre document (and its typical building blocks) ourselves — will enrich our ability to distinguish “good” GenAI output from “sub-par” GenAI work product (the new skill of output discernment). By learning how to write a genre document first without GenAI help, we enhance our understanding of how each document’s logic should “flow,” the challenges and intricacies of seamlessly threading component parts together, and the difference between generic and nuanced versions of relevant content.

After “doing one,” let’s garner feedback from multiple sources:

• Ourselves — checking the quality of our sentences, paragraphs, and document-as-a-whole against the rubric we created while “seeing one”
• Experts in our discourse community — asking mentors, teachers, professors, supervisors, colleagues, or peers to evaluate our document draft based on our (or their) rubric
• GenAI — feeding our document (or pieces of a lengthy document) into a GenAI tool and asking the chatbot to evaluate the content based on our rubric

When asking a GenAI chatbot to evaluate a “do one” draft, Professor Joel Gladd, Department Chair of Integrated Studies at College of Western Idaho, indicates that “simple inputs such as ‘leave feedback on the following draft’ will often be too open-ended.” Instead, he suggests the following prompt structure:

• Provide the chatbot with guidelines for delivering feedback, including the quality criteria for each component of the document (For an example of how to phrase writing rubric criteria, please scroll to the bottom of Professor Gladd’s book chapter.)
• Describe how the chatbot should deliver such feedback (Professor Gladd recommends instructing the chatbot: “When leaving feedback, first notice what the writer is doing well (‘I like how you…’), then point out areas that could use more development or corrections. All feedback should refer to specific paragraphs or parts of the [document]. Include specific phrases or sentences as much as possible.”)
• Give the chatbot the document draft (or pieces of a lengthy draft).

Once we are comfortable with our “doing one” skills, we can move on to “teaching one” to the GenAI chatbot.

Teach One: Instruct GenAI How to Generate High-Quality Writing

Rather than racing to our laptops and asking GenAI to churn out a genre of writing from scratch, let’s decelerate and “teach one” first. For short writing genres such as blogs or one-page essays, we likely can teach the GenAI chatbot to generate the full document in one series of well-structured prompts. But for longer, more complex genres, we should teach the GenAI chatbot how to write one component part or building block of the document at a time.

When asking GenAI to write for us (or ideally, with us), experts recommend giving the chatbot lots of context and parameters: who we are, what exactly we want, the purpose of the document, the “role” we want the chatbot to assume in drafting it, the specific task we want performed, examples of “good” output, the tone or style we prefer, and the format we desire. (When working with open/public GenAI tools like ChatGPT, Google Gemini, Claude AI, etc., we need to be mindful not to include confidential, privileged, or proprietary material in our prompts.)

Here is an example of how a writer — in particular, a legal writer — might prompt a GenAI chatbot to summarize a legal precedent case for eventual insertion into a lengthy brief to be filed with a court:

• Give the chatbot context: “I am a lawyer writing a persuasive legal brief to be submitted to a court on behalf of a party in a litigation.”
• Tell the chatbot what role you want it to play: “Please act in the role of a lawyer.”
• Instruct the chatbot on the task you want it to perform: “Please use the ‘rule explanation template’ I will provide you in a moment to summarize the legal case of [insert precedent case name] in one or two paragraphs, for incorporation into a legal brief I plan to file with the court.”
• Give the chatbot examples: “The following is a template of a good ‘rule explanation’ using a legal case to illustrate a legal rule [insert template]. Also, here is an example of a good ‘rule explanation’ paragraph [insert example].”
• Explain your quality criteria to the chatbot: “A good rule explanation paragraph includes the following component parts: (1) a transition to the case such as ‘For example’ or ‘For instance,’ (2) an accurate case citation, (3) a few sentences summarizing the case facts, (4) one sentence stating the court’s ‘holding’ or decision in the case, (5) a few sentences describing the rationale or reasoning underlying the court’s decision.”
• Tell the chatbot what tone or style of writing you want it to use: “Please use the professional tone and writing style of an ethical, diligent, and respectful lawyer.”
• Tell the chatbot what format you’d like for the result: “Please summarize the case in one or two paragraphs.”

Professor Joe Regalia recommends separating multiple prompt instructions with “delimiters” (symbols like ***) so the chatbot doesn’t get confused trying to process lengthy guidelines.

Once GenAI has generated the output (usually within a few seconds!), we can ask it to evaluate its own work against our rubric.

Using GenAI to Refine Our Work

After deploying SODOTO to teach our favorite GenAI tool how to generate a preliminary draft of a genre, we should engage GenAI further to refine and edit our/its work. In addition to the new GenAI skill of output discernment, we must learn how to competently move a GenAI-generated draft along a continuum from rough-draft to final-human-vetted form. In the fine-tuning phase, we can deploy GenAI to:

• Evaluate the document’s logic flow
• Recommend smooth transitions between paragraphs or sections
• Generate “themes” or analogies to make the prose more interesting
• Craft engaging headings to “signpost” different sections of the document for the reader
• Draft (or strengthen) thesis sentences or topic sentences within each paragraph or section
• Edit paragraphs to ensure they have a beginning, middle, and end
• Revise sentences to convert passive voice to active voice
• Change dull verbs to more compelling or vivid verbs
• Check sentences and paragraphs for punctuation or typographical errors
• Check for appropriate citation form

Like any new skill, the more we practice using GenAI, make mistakes, reflect on those mistakes, and try again, the faster we’ll improve our facility and dexterity with these new tools. Innovative technology like GenAI certainly can feel disruptive and disorienting at first, but if we embrace it with a curious spirit, GenAI just might become a reliable creative companion, a motivational daily messenger we need, and a much-needed encourager along our journey toward amplifying our writer voices.

**Professor Heidi K. Brown is Associate Dean for Upper Level Writing at New York Law School. She teaches legal writing and designs workshops, courses, and curricula around “writer identity formation,” including incorporating GenAI tools into writers’ workflow. For more, check out www.theflourishinglawyer.org or email Heidi at heidi.brown@nyls.edu.