Every article about AI for teachers uses the same three examples. Write a lesson plan, draft a rubric, generate discussion questions. Save time.
I understand why — they’re easy to explain in a paragraph. But if that’s the whole of what you’ve tried, and what came back was generic and needed fixing anyway, you probably concluded the thing was oversold. That’s a reasonable conclusion to draw from that evidence.
I’ve been using Claude nearly every day for about two years. I teach debate and AI to students ranging from elementary school through university, plus adult learners. I run summer labs, I’m finishing a book, and I put out this newsletter. Below is what I actually do with it, in roughly the order I came to it.
Two things about how I work explain most of the list.
I don’t write careful prompts. I’ve read the guides on how to do it and I don’t do it. What I type instead is the same half-legible note I’d scribble on the back of a folder — the drills we ran, who debated whom, the four things I want on the slides — and then I upload whatever materials I already have sitting on my computer and let it work out what I meant. Most of what follows is that same move applied to a different pile of paper.
And I often don’t type at all. There’s a microphone button in the message box and I use it constantly. I talk, it transcribes, it goes from there. A good share of the notes described below started as me talking rather than writing, including a fair amount of this post. I also have the app on my phone, which is the same account with the same conversations in it, so I can start something in the car and finish it on the laptop that night. That’s turned a lot of ten-minute gaps into usable time — between rounds, walking to lunch, sitting in a pickup line.
The free year
On July 14, Anthropic launched Claude for Teachers. Verified K-12 educators in the US get free access to premium Claude, a library of teaching skills, and a connection to evidence-based curricula mapped to academic standards in all 50 states. Sign up by June 30, 2027 and you get a full year.
The standards connection is the part that goes beyond a discount. Through Learning Commons, Claude has each state’s standards along with the smaller competencies underneath them and the order students usually learn them in, so what you get back is scaffolded rather than just on topic. It also reaches curricular resources like OpenSciEd and Illustrative Mathematics, and connects to nine K-12 tools directly, including Canva Education, Diffit, MagicSchool, and TeachFX.
It includes Claude Code and Cowork. I’ll come back to both.
On privacy: teacher data isn’t used for training, student information falls under a K-12 data processing addendum written to comply with FERPA, and the terms were built alongside a standard from the American Federation of Teachers. Note the line Anthropic added a week after launch, though, because it matters more than the rest of it: whether and how educational data may be used is determined by your district and state policies. The product doesn’t override your district.
Eligibility is narrower than the headlines suggest. It’s individual US K-12 educators. A schools-and-districts version is coming and districts are pointed to Claude for Nonprofits in the meantime. If you teach at a university, in adult education, or outside the US, this particular offer isn’t for you. Nothing else below depends on it.
There’s also a free AI Fluency course for PK-12 teachers built with Teach For America, Creative Commons licensed and model-agnostic, which is a good thing to walk into a department meeting with.
The plans, quickly
The pricing page confuses people because the plan names don’t tell you what changes between them.
Free gets you chat, web search, file creation, memory, and connectors, which is more than most people realize. Pro is $20 a month, or $17 if you pay annually, and adds usage, Claude Code, Cowork, Claude Design, unlimited projects, and Research. That’s the tier the teacher account is built on. Max starts at $100 and buys 5x or 20x Pro’s usage plus higher output limits and priority access. Team runs $20 to $100 per seat depending on the seat, with shared administration, and Enterprise is $20 a seat plus usage.
I pay for Max, and I want to be plain about what that buys, because I’d rather nobody read this post and conclude they need to spend money. It buys me room and one extra model. Everything described here works on the free teacher account. Usage limits reset on a rolling five-hour window, so what a lower tier really constrains is how much you can do in one sitting.
And the models
You’ll run into the names. Haiku is the fast cheap one and Sonnet is the everyday default. Opus is the heavy one, and Opus 5 is what I used for nearly everything in this post. It’s also what the teacher account gets. Fable sits above Opus and is metered tightly even if you’re paying. Above that sits Mythos, which essentially nobody has — it’s out with a small number of organizations under a program called Glasswing.
Don’t spend an hour on this. Opus 5 is excellent, you have it, and whether any of this works for you won’t turn on the model.
The list
1. Feedback on everything before students see it
Almost every lesson, handout, and set of class materials I make goes into Claude before it goes to a student.
Sometimes I just ask for general feedback. More often I ask something specific, and being specific is what makes it useful: will an eleven-year-old follow this, will this interest a college student, which words here are they not going to know, where does the pacing sag, is the third section trying to do too much.
Because I teach across such a wide range, the level question is the one I ask most. Material that lands fine with college students loses a middle schooler by the second paragraph, and it’s almost always vocabulary I stopped noticing was vocabulary sometime around 1998.
2. Building a deck from a scribble
I type a chicken-scratch outline of what I want to teach. Fragments, arrows, three words where a sentence should be, the order approximately right. Then I ask Claude to do some web research on the topic, draw on what it knows, and build the presentation.
It works out what I meant, which is the part I’d underline if I were only going to say one thing. I’m not writing a specification. I’m writing the note I’d leave for myself, and the distance between that note and a finished deck is work I simply don’t do anymore.
I do this even for topics I know cold, which took me a while to be willing to try. I could type it all out. But seeding it and letting it research and assemble is faster than typing, and what comes back is a solid first pass.
Building from scratch does need more editing than the next item on this list. It’s still ten minutes against two hours. And the more it knows about you — a textbook you wrote or teach from, articles you’ve published, decks from your own classes — the closer that first pass lands.
And, yes, it even adds the speaker noters.
3. Rewriting every example when the topic changes
I have a small textbook for each debate event. Teaching the basics works best when the examples come from the resolution students are about to argue, so they’re learning the format and the topic at the same time. But topics rotate, and for years I left the examples alone, because rewriting a whole textbook was never going to make it to the top of the list. Students got examples about a topic nobody had debated in two years. I knew it was shortchanging them and I did it anyway.
Now I hand over the new topic and it rewrites the examples throughout. It understands the topic well enough to break each example cleanly into the Pro version and the Con version. I proofread all of it. Ninety-nine percent of the time I change nothing.
Same thing with slide decks, both mine and the ones on DebateUS.org that other coaches teach from. Decks do need a few tweaks afterward, which I make by typing something like “on slide nine, change these three things.” One deck at a time works today. I think batching a whole folder is within reach with what I’ve already got set up, but I haven’t pushed on it yet.
Any material whose examples ought to track something current — an election, a case study, a text you swapped out over the summer — can be brought up to date in an afternoon rather than never.
4. Research
Turn web search on, and turn on Research for anything that needs depth.
For my purposes it beats Google, mostly because I’m not clicking through ten blue links hunting for the paragraph that matters, and because it explains what it found instead of handing me a stack and wishing me luck. It’s strongest when the question is narrow.
Last year one of my debaters needed to attack an argument built on Boltzmann brains. I’d never heard of them. I googled the garbled version I had in my head and got nothing useful. Claude explained the concept, showed me where it’s vulnerable, built the arguments against it, and gave me sources to read. My debaters say the same thing about the tight specific stuff, which is most of what they need.
5. Catching my own mistakes
People say AI is sycophantic. We did go through a stretch where that was true. It isn’t what I run into now.
Claude tells me when I’ve misunderstood something, and fairly often gives me a better version of what I was fumbling toward. While I was putting this post together I referred to that concept as “Bozeman brains” and it stopped me — you’re thinking of Boltzmann brains — and then explained them again.
Check anyway. Look it up yourself, ask a different model, click through to the sources. In my experience Claude is usually the one that turns out to be right. I don’t think that means it’s smarter than I am, but it does understand some things better than I do, which is about what you’d expect given how much is out there that I haven’t read. I find that unremarkable. I know people who find it upsetting.
6. Talking through an idea
A lot of these posts start this way, and it’s the least technical thing on the list.
I open the app, usually on my phone, hold the microphone, and start talking. No technique. “I’ve been thinking about this.” Recently it was that so much of education is now standardized — what students have to know, how it gets taught, when — while the singularity means unpredictable change and none of us knows what the world looks like on the other side. Those two things can’t both be true. A standardized curriculum assumes a destination we can see.
I lay the idea out, ask for feedback, then ask flatly for ideas I haven’t had. It pulls from education, history, sociology, fields I’ve read some of but not exhaustively, and it comes back with angles I might have gotten to eventually. Whether that makes them original is an argument I don’t need to win.
Then it’s back and forth with no structure to it at all. It offers something, that shakes something loose for me, it pushes on what I said. It’s the conversation you’d have with a colleague who happens to be free.
Most teachers won’t use this for essay ideas, though. They’ll use it for the problem in the room.
Something isn’t working and you can’t put your finger on why. The discussion dies every time you hand it over. A student stopped participating three weeks ago and you can’t tell whether to push or leave it alone. The same unit has fallen flat two years running. You’re teaching two sections of the same course and only one of them works.
Talk it through the same way. Describe what’s actually happening, including what you already tried and how the room felt. Ask what you’re missing. Then push back, because a good half of what comes back won’t fit your students and you’ll know which half immediately. Say so out loud: that won’t work with this group, they’ll take it as a punishment. It adjusts, and the second round is better than the first, partly because you just articulated something you knew but had never said.
It isn’t that it has the answer. It’s that it asks the questions a good colleague would ask, at ten at night, when there’s no good colleague around. And every so often it names the thing you’d been circling for a month — that this isn’t an engagement problem, it’s that they can’t do the prerequisite skill and are covering.
Two limits I’d state plainly. It doesn’t know your students, your building, or your department’s politics (at least until you work with it for awhile), so everything it offers is a hypothesis you test against things only you know. And it’s not a replacement for the colleague down the hall. It’s what you use before that conversation, or to work out what you want to ask.
The write-up. Once I have the ideas I want, I ask for it as a post. Claude knows my voice by now — it’s seen a lot of my writing, I’ve had it study the newsletter, I’ve fed it samples, I’ve adjusted my settings, and I’ve built skills. So I say “write this up” and it does.
Then verification, which matters more than the drafting does. I ask it to verify every claim in the piece and attach a hyperlink to something that supports it. I check the links. It’s been a couple of months since I found an error. More useful is what it tells me when it can’t find support: I can’t document this. Sometimes it just tells me I’m wrong. Then I either add the source I know exists, cut the idea, or relabel it as a conjecture I happen to hold rather than a claim I’m making. That response is a decent map of where my thinking is solid and where I’ve got more work to do.
Then revision. I read the draft in the panel on the right and type up a numbered list of changes as I go. I reread and work paragraph by paragraph — you said this, say it a little differently. Then I open it in a Google Doc, since I have Drive connected, and make the last edits by hand, though there usually aren’t many. Sometimes I paste the whole thing into ChatGPT and ask for feedback, which is very good, and bring the parts I agree with back to Claude to implement. Final read, and post.
That’s four passes before anything goes out, which I mention because “he talks to it and it writes the post” isn’t what’s happening.
7. Recording myself teach
This one is newer and I’m leaning on it more.
I record with Voice Memos whenever I’m doing anything lecture-shaped. I’m not much of a lecturer, but new debaters need catching up and sometimes that runs twenty minutes.
The point isn’t the lecture. It’s that teaching live produces material the prepared materials don’t have. I improvise an example that works better than the one on the slide. I write something on the whiteboard that isn’t in the deck at all. A student’s question pulls a long explanation out of me. I notice exactly where the room stopped following.
Voice Memos will give you a transcript if you tap it. I email it to myself, drop it into Claude with whatever I taught from that day, and say: update the materials based on what I added as I went. That’s the entire prompt.
When there’s time I also record about two minutes at the end of class while it’s fresh. What went well, what didn’t. I had to walk them through that much more slowly than I planned. I used vocabulary they didn’t have. I did more checking for understanding than explaining. That catches teaching quality and not just missing content, so the revisions improve how I teach the thing rather than only what’s in it.
This closes a loop back to the first item, incidentally. There I’m predicting what students will find hard. Here I find out. Run that a few cycles and it starts to know what your students struggle with rather than what’s generically difficult.
One caution. Recording in a classroom picks up student voices, and plenty of districts have rules about that. Recording only your own presentation stretches keeps most of the value and avoids most of the problem. Check your policy before you check mine.
8. Class reports
My online classes are recorded as a matter of program policy. I run Fireflies on top of it, though a plain Zoom transcript would work fine.
The transcript goes in with a prompt as short as “generate a class report,” or “for parents and students” if they’re young. What comes back covers everything we did and reliably foregrounds the things I most wanted them to hold onto. It figures out on its own that there was homework and puts it under a homework heading. It tracks that we’ve got a practice debate coming and who’s on which side.
One report added a section, unprompted, on how parents can help their kid become a better debater. Ask what they’re arguing about this week. Here’s roughly how the format works. Quiz them at dinner. It would not have occurred to me to ask for that, and every one of those parents wants to help and has no idea how.
9. Daily lab reports at camp
Camp is where this stops being a convenience. Six or seven hours a day, and by Thursday nobody remembers what happened Monday. Families paying for an intensive see almost none of it.
I type a chicken-scratch outline of what we actually did — the drills, who debated whom, what came up — and upload all the instructional material we used that day. Then I ask for a comprehensive report for parents.
The upload is what makes it work. My outline is a skeleton. The materials are everything else. It can see the morning block was on frontlining because the deck says so, and it can explain frontlining to a parent because the deck already explains it to a student.
Two rules keep it honest. Students get consistent pseudonyms and coaches keep their real names, so the report can circulate without putting a fourteen-year-old’s worst speech in front of strangers. And when my outline gives the shape of the day without the content of the speeches, the report says so rather than inventing them. An honest gap beats a confident guess, which is what I ask my debaters to do with evidence, so a report that broke the rule while teaching it wouldn’t be worth much.
10. Individual student evaluations
Twice a session, so once every four days, each student gets a written evaluation on the program’s own form. Twenty students, fixed rubric, hard deadline. This used to eat an entire evening and produce twenty paragraphs that all sounded the same by the end.
Setup is a one-time cost. I gave it the PDF form and had it learn the evaluation categories, so everything comes out keyed to the rubric instead of to whatever I felt like writing.
After that it’s collection rather than writing. I type a line about a student when something’s fresh. Near the end of a block I hand my phone around and have each student record their name and one thing they learned, which takes ninety seconds for the whole room. My co-instructor sends me a voice file with his comments on everybody, which he can record walking to lunch instead of typing at midnight. All of it goes in at once, with instructions to write a report on each student drawing on both the coaches’ comments and the student’s own, fill out the form, and save each one as its own named PDF.
That last part sounds like nothing. Twenty correctly named files turns distribution into a five-minute job instead of an hour of clicking.
What makes these better than what I used to write is the student audio. A report that reflects a student’s own account of what they learned tells a parent something a rubric score can’t, and it tells the student that the thing they said out loud got written down somewhere.
11. Vocabulary lists
The smallest thing here. Upload the lesson, ask for every term a student would need to follow it, defined at the level of the room.
It catches what I’ve stopped seeing. Warrant, impact, turn, clash — none of those register as vocabulary to me anymore. Students hit an unfamiliar word and a familiar word being used strangely and can’t tell the difference. Ask for the list at the level you’re teaching, and ask specifically which terms you use without ever defining.
12. Making the documents look like something
A report that shows up as a wall of plain text gets skimmed. The same content with a title block, real typography, and a clear structure gets read, and parents keep it.
I finish the content first and then ask for it as a designed document. A serif face for reading, monospace for labels and dates, color separating the Pro material from the Con, callout boxes for coach notes. It writes the HTML and the print stylesheet and renders a PDF. Then I revise off the rendered pages: page six is nearly empty, tighten it to five. Move homework above the schedule. That accent color is too loud. I never open a design program.
When I want images in it, I push the content into Gamma through the connector, which themes it, generates illustrations, and lays it out for further editing on their side.
The design is carrying information, which is why I bother. When the two sides of an argument sit in visually distinct panels, a parent who has never watched a debate can tell that the event has sides and their kid was assigned one, before they’ve read a word.
13. Long documents that have to hold up
A newsletter post is one thing. A book chapter with a hundred endnotes is another, and so is anything a school publishes under its own name.
I’ve been writing a book this way. Every chapter goes through the same sequence: draft, verify every claim against a live source, convert to Word with the notes collected as endnotes, then a pass that strips out the words I’ve told Claude I don’t write with. The mechanics involve a document conversion tool you’ve never heard of and don’t need.
Three habits underneath that do transfer.
Write your house style down once. How headings work, how sources appear, what reading level, what tense the objectives are in. A department handbook written by four people over two years has four voices in it. One standing instruction fixes every document that comes after.
Keep a vocabulary blacklist. Terminology your district uses, terminology it dropped, phrases that mean one thing in your building and something else outside it.
And keep verification as its own step. Not “write this and make it accurate,” but write it, then go back and check each claim against a source, then tell me what you couldn’t support. Splitting them apart is what makes the second one honest.
14. Materials other people teach from
A slide I made for myself can be thin, because I’m standing next to it filling in what it doesn’t say. A slide another coach teaches from has to carry its own weight.
The textbooks, the argument decks NAVI coaches use, the Congress dockets, the Value of Debate materials, an AI literacy curriculum I built with Anand Rao for Oxford University Press, research briefs on hegemony and on women’s rights against sovereignty claims. When I say the audience is a first-year coach rather than a veteran, the materials come back with scaffolding I’d have left out on my own.
Curriculum materials don’t go stale because nobody noticed. They go stale because fixing them costs more hours than anyone has, and that particular constraint is mostly gone now.
15. Building software when you can’t write software
I’ll be careful here, because this is the item most likely to read as exaggeration.
My coding abilities are extremely limited. And I built a Chrome extension that checks the claims in an article while I’m reading it. I open a piece and it flags what’s being asserted without support.
I built it by describing what I wanted in plain English and answering questions when it asked them. That’s the whole method. The tool is Claude Code, which comes with the teacher account and every paid plan. People describe it as a developer tool, which is fair and also incomplete, since I’m not a developer and it worked anyway. The going term for this is vibe coding. It sounds unserious and it’s accurate.
I’m using it to rebuild DebateUS.org too, which is further along than I expected to get.
I’m not suggesting every teacher should build software. The narrower point is that the thing you’ve been assuming is out of reach — the script that sorts that folder of submissions, the little piece of classroom software that doesn’t exist because the market for it is one person — might not be out of reach anymore. The barrier used to be a language you’d have to spend a year learning. Now it’s being able to describe what you want the thing to do, and teachers are unusually good at that.
Start smaller than a Chrome extension
Don’t start with Claude Code. Start in the chat window, which will build you a working interactive thing right there.
Say you’re reviewing a unit. Instead of asking for twenty review questions, ask for the game: build me a review game for this unit, two teams, questions get harder as they go, a timer, a running score, and when someone’s wrong it explains why instead of just buzzing. Then upload the unit materials so the questions come from what you taught rather than from nowhere in particular.
What comes back shows up beside the conversation and runs. You play it. Then you fix it the way you’d fix anything else — the timer’s too short, round three is too easy, cut the wrong-answer explanations down, add a round where they have to justify the answer out loud. Every one of those is a sentence.
That same move gets you most of what people pay for classroom software. A self-scoring quiz that gives real feedback per wrong answer. A drag-and-drop sorter, so put these claims under the right argument or these events on the timeline. A matching or flashcard drill built from the vocabulary list in item 11. A simulation where students choose, see the consequence, and choose again. A generator for practice problems or discussion prompts or debate matchups. A scoreboard for a class tournament.
You can share what you build with students by link, so it isn’t only a projector activity.
You’d move up to Claude Code when the thing needs to hold data over time, run outside a browser tab, or live permanently on your site. That’s where I ended up with the extension and the website. Almost nothing a classroom needs starts there, and starting there is why most teachers never start.
The three places context belongs
Most people put everything into every message and then wonder why the output drifts.
Settings are for who you are. A standing block that applies to every conversation you have. Mine is short and mostly negative — be concise, don’t pad. If you catch yourself typing the same correction over and over, that correction belongs in settings. Connectors live here too. I have Drive connected, which is why drafts open as documents instead of text I have to copy somewhere, and Gamma connected, which is what makes the design handoff a single instruction. The teacher account adds Learning Commons and the K-12 tools.
Projects are for what you’re working on. A walled workspace with its own uploaded knowledge and its own instructions, where every conversation starts from those. For a teacher the natural unit is the course: syllabus, the textbook you actually teach from, unit plans, department style rules, last year’s materials. Then “write me a warm-up for Thursday” already knows what Tuesday covered. The way this goes wrong is one enormous project called Teaching, where context that applies to everything ends up applying to nothing.
Skills are for how a document gets built. Standing instructions that load when the task calls for them: the format, what has to appear and in what order, the audience, the words you won’t use.
You don’t write one from scratch, which is the part people miss. You build it out of a conversation you already had. Work an artifact until it’s right — for a deck that meant twelve slides instead of twenty-five, every concept illustrated from the current resolution, every example split into both sides, nothing over about thirty words a slide — and then say: take everything we just worked out and write it up as a skill. The next deck arrives built that way.
Mine are the recurring documents of my working life. Congress dockets, verified pro-and-con research briefs, argument explainers in a coaching register, extemp briefs, one that writes in the NSDA’s public voice and a separate one for the register a board uses with itself, an AGI forecast tracker, and the class report. There’s also one for researching home sales in New Jersey, which should tell you the bar is lower than you think.
A skill guarantees the report is complete and every so often costs me the section I never would have thought to ask for, so I run some jobs without it on purpose. If you’d rather start from something that works, Anthropic open-sourced its own K-12 teaching skills at github.com/anthropics/k12-teacher-skills, and they’re worth reading even if you never install one.
Cowork is the same Claude pointed at a folder instead of a message box. You describe a job, it plans, it works through the files, and it reports back. Twenty student evaluations saved as twenty named PDFs. Eight decks that all need the same topic swapped through them. A week of reports in one design. In chat you supervise each of those one at a time; in Cowork you describe the job once. It started as a desktop app, and desktop is still the fullest version since that’s where it reaches files on your own machine, but it runs on the web and your phone now, sessions keep going remotely, and you can start something at your desk and pick it up later with the laptop shut.
Where I’d start
Don’t try fifteen things. Try two.
Take the next thing you were going to write anyway — a handout, a study guide, next week’s slides — and ask for feedback on it at the level of the students who’ll actually get it. That’s the first item on this list, it takes five minutes, and it’ll tell you more than this whole post did.
Then the next time you’re teaching and you hear yourself say something that isn’t in your materials, record the rest of the class (if legally permitted) and feed the transcript back with the deck. That’s item seven, and it’s the one nobody thinks of, because we’ve all quietly accepted that the best explanation we ever gave was the one nobody wrote down.
The rest is variations on those two. And if you take one thing from the whole list, take this: you don’t have to write a good prompt. I never have. I type the notes I was going to type anyway, hand over the materials I already made, and let it work out what I meant.