Fixing next period's slides while printing tomorrow's worksheet, then spending lunch scanning submissions — most teaching days are "making" and "checking" stacked back-to-back, and "returning work to students" barely fits before the bell rings.
I've spoken with teachers who tell me some version of the same thing: "I tried Claude, but I can't quite pin down where it actually helps." A tool can look promising in the abstract and still fail to stick in your weekly rhythm if you don't have a concrete model for when to reach for it. That model lives in three pillars: lesson prep, grading, and personalized feedback.
This article walks through practical patterns that work in public schools, private schools, cram schools, and universities alike. It isn't just "what prompt to send" — it also spells out where you shouldn't use Claude at all.
Why Claude fits classroom work (and where it doesn't)
Generative AI has real strengths and real blind spots. Before the specific use cases, three framing points are worth holding in mind.
First, Claude is excellent at structured long-form writing. Lesson plans, unit overviews, rubrics — most teacher deliverables follow a known skeleton (objective, assessment criteria, time allocation). Claude is strong at filling in that kind of scaffold quickly and consistently.
Second, there's a clean line between draft-quality evaluation and final grading decisions. Claude can help you triage student responses and spot patterns in common misconceptions, but handing it a rubric and asking for a letter grade is a different thing entirely — grading policy is school-specific and you need to own that decision to defend it later.
Third, student personal data is off-limits. Names, student IDs, family situations, and health information should never go into a prompt, even on a free plan. The privacy section at the end covers the practical rules.
Lesson prep: turn unit plans and worksheets into a template
Lesson plans and worksheets have fixed slots — objectives, activities, assessment criteria — so once you lock down a format, you can produce them surprisingly fast. A prompt like the one below returns a usable first draft in under a minute:
You are an experienced middle school science teacher.
Create a single 50-minute lesson plan given the following:
- Grade: 8th grade (age 13–14)
- Unit: Current and voltage (Ohm's law)
- Objective: Students can calculate current in simple circuits using Ohm's law
- Class context: 50-min period, 35 students, lab access available
- Output format:
1. Today's objective (2 sentences)
2. Activity flow (5-min opening, 30-min main, 10-min wrap)
3. Assessment criteria across three dimensions: knowledge, thinking, engagement
4. Predicted student stumbling points + how to respond verbally
The non-obvious move here is the fourth output slot. Everything above it is structural filler that any decent model can produce. The prediction of where students will get stuck is where teacher expertise actually lives, and you have to explicitly ask for it — otherwise Claude will happily give you a clean-but-generic plan that skips the part you actually need.
Worksheets work the same way. Specify a distribution like "5 basic / 3 applied / 1 stretch problem" and you get tiered practice sets in a few minutes. One rule: solve every question yourself before you hand it out. Claude's math questions occasionally have phrasing quirks that can confuse a 13-year-old trying to parse them at speed.
See the complete guide to Claude for education for use cases beyond lesson prep alone.
Grading drafts: use Claude for "direction checks," not scores
Grading short-answer and essay questions is one of the heaviest time sinks in the job. That said, I wouldn't hand the grading itself to Claude. Scoring bands are tied to school-specific policy, and if you can't defend a grade in your own words, you shouldn't be assigning it.
What works well is narrower: a direction check against the rubric. You give Claude the criteria and ask it to sort the response into three buckets — "clearly met," "clearly short," and "judgment call" — without ever producing a number.
Against the rubric below, sort Student A's response into three categories:
"clearly met," "clearly short of the criterion," and "judgment call."
Do not output a score — I'll decide that.
Rubric:
- Point 1: Addresses both human and natural causes of climate change
- Point 2: Cites at least two specific examples
- Point 3: Links personal view to supporting evidence logically
Student response:
(Paste body here. Strip name and student ID first.)
Now instead of sweeping every rubric point on every response, you focus on the judgment calls. Teachers I've talked to have cut essay-grading time roughly in half this way. Before you paste a real response, though, read the privacy section below — it's not optional.
Personalized feedback: generate drafts, not finals
Parent conferences, end-of-term comments, feedback on club reflections — any situation where you need to write thirty to forty short, personal notes in a day is where Claude's drafting speed pays off most.
The trick is to feed concrete facts, not abstractions. If you hand over vague descriptions, Claude hands back generic text that could fit any student. Four specific observations produce something that actually sounds like it's about this kid:
Write a ~150-word personalized comment on the submitted report for the
anonymous student described below. Use a warm tone and include one specific
piece of advice for next time.
- Fact 1: Submitted 3 days before the deadline
- Fact 2: Cited 4 references
- Fact 3: The conclusion drifts slightly from the main argument
- Fact 4: Has been more vocal in class this past month
Treat the output as a draft — you rewrite the final in your own voice. The judgment that says "don't mention club activities right now because they're going through something" is the part no model can handle for you. That's precisely why teacher-in-the-loop isn't optional here.
If you want to tighten the prompts above further, the system prompt design guide covers the underlying principles.
Student privacy: the rules you set before you start
This is the gate. Get it wrong once and it becomes a parent-facing and administrator-facing problem very quickly. A few non-negotiable rules:
- No names, student IDs, or school names in prompts — use "Student A," "8th grade class."
- No family situation, health, or developmental records. These belong in your school's internal systems, not in an external AI prompt.
- Strip identifying info before pasting responses. The name field at the top of an essay needs to come off first.
- Check your school or district's AI policy. If there's an official guideline from your IT office or the board of education, that takes precedence over anything here.
- Consider the Claude academic plan if you qualify — it gives faculty and students discounted access to stronger models, subject to an institutional email check.
Decide your rules once, share them with your department head or vice principal, and stick to them. The teachers who get into trouble with AI tools are usually the ones who started using them before deciding the policy.
A realistic starter checklist
You don't need to overhaul your whole workflow in a week. Pick one of these and stop there for the first seven days:
- Choose one single use case to try (direction checks on grading, say).
- Save one reusable prompt as a template.
- Read one page of your school's AI policy.
- Show the output to one colleague and ask what felt off.
Tools like Claude stick when they match the way you already work, not when they force you to reinvent your week. By the end of that first week, you'll have a personal list of "this works" and "this doesn't" — and that list is what turns an interesting demo into something you actually use on a Tuesday in March.