The teacher remains the instructional designer
ChatGPT is good at producing options. It can turn a learning objective into an activity sequence, suggest examples, rewrite directions at a different reading level, or give me five ways to check understanding. It does not know whether my ninth-period class needs movement, whether yesterday's misconception is still present, or whether a suggested example will land badly in my community. That is the part I bring.
I get the best results when I give the model the decisions I have already made: the objective, time available, materials, prerequisite knowledge, reading range, accessibility supports, and the evidence I want to see. I ask for a draft that exposes assumptions and missing information. I then make the professional choices about what belongs in the room.
The US Department of Education's guidance treats AI as something that should support, rather than displace, educators and learners. That principle is practical. When a generated lesson saves me twenty minutes, I can spend those twenty minutes improving the example, planning a better question, or thinking about a student who needs a different entry point.
A Monday planning workflow I would actually use
I would not begin with “make me a lesson plan.” I would start with the standard or objective and write down what students should be able to do by the end. Then I would add the length of the period, available materials, prior knowledge, likely misconception, and any support I can realistically provide. Those constraints keep the answer connected to my classroom instead of a fictional ideal classroom.
Next I ask for a sequence with teacher moves, student actions, checks for understanding, and an exit ticket. I want the model to show me where I will learn whether students are following, not just fill the page with activities. I ask it to offer one simpler entry point and one extension while preserving the same objective.
Finally, I edit against my curriculum and my students. I remove activities that require technology we do not have. I correct invented standards references and examples. I make the language sound like me. The saved version is the reviewed lesson in my planning system, not the chat transcript.
Differentiation that preserves the learning goal
“Make this easier” is a weak request because easier can mean shorter, less rigorous, or simply less interesting. I ask ChatGPT to hold the objective constant and vary the support. For a reading task, that may mean a vocabulary preview, chunked text, sentence frames, or an audio-friendly version. For a math task, it may mean worked examples, a visual representation, or a gradual release of hints. For an advanced learner, it may mean a new constraint or a transfer problem.
I review every version for dignity and access. A scaffold should not announce that a student is receiving the “easy” work. It should not remove the thinking I am trying to teach. I also check whether the suggested support is possible with my staffing, time, and materials. AI often proposes a beautiful three-station lesson when I have one teacher, twenty-eight students, and forty-five minutes.
A useful test is to ask: can I explain how each version produces evidence of the same objective? If I cannot, the prompt needs revision. This turns AI from a worksheet machine into a way to think more deliberately about access.
Feedback is more useful than automatic grading
I am comfortable asking ChatGPT to help me phrase feedback. I am not comfortable making it the final judge of a student's work. A rubric is not a magic spell: the model can misread evidence, reward a polished style over an original idea, penalize a language learner, or invent a weakness because it expects one to be there.
My safer workflow is to provide a de-identified sample and the rubric, then request one specific strength, one priority revision, and one question that returns the thinking to the student. I ask the model to quote the evidence it used and to say what it cannot evaluate. I decide whether the feedback is fair, useful, and appropriate for this learner before it reaches the student.
I also use process evidence. Drafts, revision history, conferences, oral explanations, and in-class writing tell me more than a detector score. AI detection can be a prompt for a conversation, but it should not become a shortcut to a disciplinary conclusion. The student's opportunity to explain their work matters.
Parent messages, progress notes, and the human tone
Family communication is a good drafting use because the hard part is often getting a clear, calm first version onto the screen. I can give ChatGPT a few factual notes and ask for a message that explains what happened, what we are doing next, and how the family can respond. I specifically tell it not to diagnose, blame, promise an outcome, or add details I did not provide.
Then I rewrite it. A message to a family is not a generic customer-service email. It has a relationship, a history, and a tone that the model cannot see. I check names, pronouns, dates, accommodations, and the difference between an observation and an interpretation. I remove anything that would be surprising or harmful if forwarded.
The same rule applies to report-card comments and progress notes. ChatGPT can help me avoid repetitive phrasing, but the comment should still reflect what I have actually observed and what the student can do next.
Substitute plans and the invisible administrative load
One of my favorite uses is turning rough notes into a substitute-ready plan. The prompt should ask for the schedule, objective, materials, directions, expected student product, early-finisher option, and what the substitute should record. It should not receive student medical information, behavior files, or a list of private circumstances.
ChatGPT can also help turn a unit folder into a checklist, create a vocabulary review, draft a meeting agenda, or produce questions for a department planning session. These tasks are valuable because they are repetitive but still need a person to check. I save the final version where the team already works so the classroom does not depend on one teacher's chat history.
For school leaders, this is where a small pilot becomes measurable. Track planning time, revision time, missing details in substitute plans, and teacher satisfaction. Do not measure success by how many prompts were run. Measure whether the work is clearer and whether teachers have more time for students.
The student privacy checkpoint
Before using an AI tool with school information, I check the district's approved-tool list and ask who controls the data, what the vendor stores, whether it uses content for training, how deletion works, and whether the district has the required agreement. The Student Privacy Policy Office advises educators to consult their school or district before using online services that collect personally identifiable information.
In practical terms, I keep names, student IDs, grades linked to names, disability or health details, family information, disciplinary records, and identifiable student work out of unapproved tools. I use placeholders such as “Student A,” remove unnecessary details, and work from an approved source when the task is sensitive. Anonymizing data reduces risk, but it does not replace district approval.
Student privacy is only one part of the review. I also consider bias, accessibility, copyright, age restrictions, account security, and whether students understand when AI is involved. A classroom rule should be visible and teachable, not buried in a vendor's terms.
Five prompts I would keep in my teacher toolkit
1. Plan from the objective
I teach [grade/course]. My learning objective is [objective]. I have [minutes], [materials], and [class constraints]. Draft a lesson sequence with an opening, modeling, guided practice, independent practice, two checks for understanding, and an exit ticket. Label what I should verify rather than inventing facts or standards.
2. Differentiate without lowering the goal
Keep this learning objective unchanged: [objective]. Create a scaffolded version, an on-level version, and an extension version of the task below. For each, name the support, the likely misconception, and the evidence that shows learning. Do not make the scaffolded task a different or less meaningful objective.
3. Build better feedback
Here is my rubric and a de-identified sample of student work. Draft feedback in three parts: one specific strength, one priority for revision, and one question that makes the student think. Quote only from the sample. Do not assign a final grade and flag anything you cannot evaluate from the evidence.
4. Prepare a substitute plan
Turn these teacher notes into a substitute plan for [grade/course]. Include the objective, schedule, materials, directions, expected student product, what to do if students finish early, and what to record for me. Use plain language. Do not include student names or private records.
5. Draft a family message
Draft a warm, factual message to a student's family about [topic] using only the notes below. Avoid diagnosis, blame, promises, and unnecessary private details. End with one clear next step and one question inviting the family's perspective. Give me a version in plain language that I can edit before sending.
What I would keep out of ChatGPT
- Identifiable student records, grades, health details, disability information, or family circumstances.
- Final decisions about grades, discipline, placement, special education, or student safety.
- Unverified facts, invented citations, or generated examples presented as classroom truth.
- Private staff, parent, or colleague information that the school has not approved for the tool.
- A complete lesson written without the teacher checking standards, accessibility, culture, and feasibility.
- A detector score used as the only evidence that a student violated an academic-integrity rule.
A 30-day classroom pilot
During week one, I would choose one low-risk task: exit tickets, lesson variations, or substitute plans. I would record how long the normal task takes and what a good result must contain. During week two, I would run the prompt on three examples and save the original, draft, edits, and final version.
During week three, I would improve the prompt based on the edits. I would add a rule for missing information, a privacy reminder, and a review checklist. If the tool keeps making the same mistake, I would narrow the job rather than asking for a longer prompt.
During week four, I would decide whether the workflow earned a place in my routine. I would compare time saved with correction time and ask whether the final material is better for students. A successful pilot should leave behind a small, understandable process that another teacher could inspect and adapt.
My bottom line
ChatGPT is most useful to teachers when it removes blank-page work and gives us more options to think about. It is not a substitute for knowing the class, understanding the curriculum, noticing a student's expression, or making a careful professional decision.
I would start with one real task, use the least sensitive input possible, ask the model to show its assumptions, and keep a human approval step. That approach is modest, but it produces something more valuable than a pile of clever prompts: a classroom workflow that saves time without surrendering responsibility.