How to Use AI for Lesson Planning, a Week at a Time

You have a standard, a reading and a Sunday night.

Somewhere between the three, a week of lessons has to appear, with a hook, an activity and an exit ticket for every day. Here is how to use AI for lesson planning to get there: upload the reading, describe the real week, ask for a skeleton before any lessons, write one day per message, check the quotes and the minutes yourself, and have a second model grade the exit tickets.

Don't ask AI for a lesson plan. Hand it your reading and your week, then make it prove every day against the text.

Skip that and you get what a substitute finds on the desk. A plan that is tidy, generic and written for a room that doesn't exist.

The example below is one full week. The standard is CCSS RI.8.2, which asks eighth graders to determine a central idea, analyze how it develops, and write an objective summary. The text is the Gettysburg Address, and the Library of Congress keeps a good reason on its exhibit page. Edward Everett spoke for two hours before Lincoln, then wrote to him the next day that he wished he had come "as near to the central idea of the occasion, in two hours, as you did in two minutes."

Swap in your own. A science standard and a lab write-up works the same way. So does a history standard and a primary source.

Here is the number that makes your Sunday worth protecting. Gallup surveyed 2,232 public school teachers, and those who use AI at least weekly estimated they save 5.9 hours a week.

That is the teachers' own estimate, and it measures their time, not their students' learning. Hours only come back if the plan survives Monday morning.

The replies quoted below come from a run of these exact prompts. Some of them broke partway through. The breaks stay in, because you will hit them too and the fix is part of the method.

How to use AI for lesson planning starts with your reading

Most teachers type a topic and a grade. "Plan a week on the Gettysburg Address for eighth grade."

What comes back is built from what the model remembers about lesson plans and about that speech. That is why it reads like every plan you have ever seen.

Attach the reading instead. In Magai, a document uploaded into the thread is the material the model works from, so the plan comes from your handout and not from memory. Do the same for the standard: paste its code and its exact wording, so the model isn't guessing which version of the skill you mean.

Say which text, too. The Library of Congress counts five known manuscript copies of the Gettysburg Address, and your class may be reading an excerpt, an adapted version or an anthology with glosses in the margin. The version in your students' hands is the one the lessons have to work with.

A warning about this example. A famous speech is easy mode, because the model knows it by heart. Your adapted passage, your district text or your own lab write-up is the real test, and there the model knows only what is in the file.

Here is the first message, as typed:

I teach 8th grade ELA. The reading is attached and the standard I'm teaching this week is below.

STANDARD: [paste the code and the full wording]

Before we plan anything, tell me in plain words what a student has to be able to DO with this reading to meet this standard. Then point to the three places in the reading where the central idea is built or turns, quoting the exact words from the text for each.

The order is deliberate. Asking for the skill in plain words first shows you whether the model read the standard the way you do, before it has written a word you would have to throw away.

The reply opened with exactly that: identify the central idea, explain how Lincoln builds it from the founding ideals to a call for action, and summarize without opinion. That is RI.8.2 in a sentence a student could repeat, so it stayed.

Then it started on the three places in the text. The first was the opening line about a nation "conceived in Liberty." The second was the turn in the third paragraph, where Lincoln says the living cannot dedicate the ground because the soldiers already have. Both quotes matched the file word for word.

The reply stopped inside the second one. The third place never arrived.

A reply that ends mid-sentence is unfinished work, and you don't plan from it. Type "Continue from the second place" and wait for the rest before you move on.

Tell it the week you actually have

A plan that ignores your Tuesday isn't a plan. The model can't see your schedule, your roster or your supply closet, so you say all of it, in one message:

Now plan the week. I have five 50 minute periods, Monday to Friday, 28 students, about six of them reading well below grade level and two English learners. No devices on Tuesday and Thursday. Give me one line per day: the learning target in student words, and the one thing students will produce that day. Friday has to end with a short written objective summary of the speech. Do not write the lessons yet.

Each sentence does a job. The counts and the period length stop it padding. The groups tell it who needs support. The device rule is a hard limit it has to build around. Friday's summary fixes where the week must land.

The last sentence holds back the lessons, because five lines are cheap to fix and five days are not.

The reply was a skeleton you can read in a minute. Here it is as a table, since five days with a target and a product each is a comparison you will scan across.

Read it the way you would read a colleague's plan. Four checks, in this order:

1. Does every target start with something a student can do? All five began with "I can" and a verb: identify, state, trace, explain, summarize.
2. Does every day produce something you can collect? Here that was an annotated copy, a statement with a quotation, an organizer, a written explanation and a summary. Nothing was "discuss," which is a product you can't grade.
3. Do the days map to the standard's verbs? RI.8.2 asks students to determine a central idea, analyze its development and its relationship to supporting ideas, and provide an objective summary. Tuesday is determine, Wednesday and Thursday are analyze, Friday is the summary, and Monday is the on-ramp.
4. Could two days swap places without a student noticing? If yes, they are one day.

The fourth check caught something. Wednesday and Thursday blur. "Trace how the idea develops" and "connect the supporting ideas to the central idea" are close to one skill, and a thirteen year old won't feel the difference between them.

That is a ten second fix at this stage:

Wednesday and Thursday sound the same to a student. Make Wednesday about how the speech moves from the dead to the living, and make Thursday about which supporting ideas Lincoln picked and what each one proves.

Had the model written five full days first, you would have found the overlap after reading all five.

Write one day at a time

With the skeleton settled, the run asked for all five days in one message:

Good. Now write Monday through Friday in full. For each day give me: a hook (under 5 minutes), the main activity with minute by minute timing that adds up to 50, and an exit ticket of no more than three questions. Use the reading's actual words wherever students need to work with the text.

Monday came back whole. The reply stopped partway through Tuesday.

Monday was good. The hook took three minutes and asked for one sentence on a question every student can answer with no background: what idea or principle should a country be built on? That is the speech's subject, reached through the students' own opinions.

The main activity moved from a first read, to pairs putting the opening sentence in plain language, to annotating the second paragraph for what the war is testing, to a class chart headed "Founded on" and "Being tested by." The target got a physical product, and the exit ticket sat in the last five minutes.

Monday also carried a support block: the text divided into paragraphs, a short glossary, and a sentence frame. "Lincoln says the nation was founded on ___, and the war is testing whether ___." The frame is usable as typed.

The block was one block for everyone, though. Your six struggling readers and your two English learners need different things, so ask for them separately:

Give me the Monday supports for the six students reading well below grade level. Then give me separate supports for the two English learners.

Five days at minute by minute detail is a long reply, and long replies hit a ceiling. The fix is to ask for one day per message:

Write Tuesday only, in full: a hook under 5 minutes, the main activity with minute by minute timing that adds up to exactly 50 minutes, and an exit ticket of no more than three questions. Respect the no devices rule. Use the reading's actual words.

Tuesday came through whole. The materials line listed printed copies, pencils and paper, so the device rule held.

The block from minute 25 to minute 34 had the teacher model the difference between a topic ("the Civil War") and a central idea (what Lincoln says about the nation and its future). That is the mistake you already know is coming, handled before it happens.

Tuesday's exit ticket was "Turn in the final central-idea statement and its supporting quotation." That is the day's product handed in, not a ticket. Hold that thought for the audit below.

Friday also came through whole. Wednesday and Thursday cut off again, even asked one at a time. Splitting a day into two messages, the activity first and the ticket second, got the tickets through but left the activities stopped short.

If a reply cuts off twice, don't ask a third time. Switch the model for that message. Type /GPT or /Gemini at the start and say "Finish Wednesday from where the last reply stopped." The thread carries your reading, your standard and the plan with it, which is the point of switching between models in one workflow. The new model starts where the last one quit.

Check the quotes and add the minutes yourself

Two checks take about five minutes and catch what the model cannot see in its own work.

The first is the quotes. Wherever a day quotes the reading, put the line next to your file. The run's hooks quoted "government of the people, by the people, for the people," then "The world will little note, nor long remember what we say here, but it can never forget what they did here," then "that these dead shall not have died in vain." All three matched the text.

That is what you would expect from a speech everyone knows. With an adapted handout, this check is the only thing between you and a model that puts quotation marks around its own paraphrase, which you then print for twenty-eight students.

You can make the check faster by making the model show where each quote lives:

For every line you quoted from the reading in this day, tell me which paragraph it comes from and give me the sentence before it.

A quote the model can't place is a quote to cut. A quote it places in the wrong paragraph tells you it was working from memory, and the rest of that day deserves a harder look.

The second is the minutes. Add the blocks yourself. Monday's came to 3, 5, 7, 9, 13, 8 and 5, which is 50. Tuesday's came to 50. Friday's nine blocks came to 50 too.

Then Friday's reply closed with its own line: "3 + 42 + 5 = 50." That describes a schedule it didn't write. The hook was 3 minutes, the closing check was 3, and the middle was 44. The total was right and the breakdown was wrong, so a stated total is a claim and not a sum.

Run the check in this order for every day before it goes in your planner:

1. Put your file in one window and the day in another, and find each quoted line in the file.
2. Add the minute blocks yourself and compare the sum to your period.
3. Read the materials line against your off limits rule, so a no devices day lists only paper.
4. Look for separate supports for each group you named.
5. Read the hook aloud and time it against the five minutes you allowed.

The table below shows what the run turned up.

Ask a second model to grade the exit tickets

The exit ticket is where plans go soft. It sits in the last five minutes of a block the model has already filled, and it gets written like a quiz.

The run shows it. Wednesday and Thursday were asked for tickets of no more than three questions, each answerable only by someone who read the speech. Wednesday's came back as: what founding ideals does Lincoln name in the opening sentence, what does Lincoln say the Civil War is testing, and what does Lincoln call on "us the living" to do. Thursday's were the same kind. What did the soldiers' actions do to the ground? What should the living take from their sacrifice? What does Lincoln hope will not perish?

Every one is recall. A student who skimmed the first paragraph passes. None asks how the central idea develops, which is what RI.8.2 asks and what the whole week was built around.

The model read "answerable only by someone who read the speech" as "a look up question," and wrote six of them.

So you audit them. Send this:

Audit the five exit tickets you just wrote against the standard. For each question, tell me whether a student could answer it correctly without having read the speech, and whether it measures the central idea and its development or only recall. Rewrite any question that fails either test.

Send it to a different model than the one that wrote the days. If /Claude wrote them, type /GPT at the start of the message, and the reverse. A second model didn't write the tickets and has nothing invested in them. In Magai the thread goes with it, so your reading, your standard and the week are already in front of it.

The audit reply in the run said it could see only Monday's ticket, because the plan had been cut off, and declined to grade four tickets it had not been shown. Then it offered proposed tickets for the other days.

That is the right behavior from an assistant. It is also why you write one day at a time first: you can't audit what never arrived.

Monday's audit was blunt. The ticket asked students to name two ideas, which was mostly recall, and it could be answered from prior knowledge of the period. The rewrite asked how Lincoln connects "conceived in Liberty" and "all men are created equal" to the war's test of whether the nation "can long endure," using both phrases in the answer. That rewrite asks for a relationship between two ideas, which is exactly what the standard asks.

The proposed tickets for the other days needed their own audit. They were better than the first drafts, and two still had problems.

Give every question one test: could a student who skipped the reading answer this? If yes, rewrite it. For a famous speech that test is harsh, because many students know something about it already. That is a reason to run it, not a reason to relax it.

Keep the parts only you can do

The model doesn't know your room. It doesn't know that third period runs six minutes late, that Wednesday has a fire drill, or that two of your pairs can't work together. Read each day aloud once, the way you will say it, and cut whatever you wouldn't.

You might say it is faster to find a ready made plan and tweak it. Fair, if it matches your standard and your text. It rarely matches both, and tweaking someone else's week costs more than describing your own.

There is one case where all of this is more than you need. If you teach the same unit with the same text for the eighth year, your plan already exists. The method is for the week that doesn't have one yet.

Keep student data out. Take names, and anything else that identifies a child, out of everything you attach or paste. FERPA is the federal law on student education records, and your district may have a stricter rule about which tools you can use at all. Check the policy first, and the compliance checklist for AI tools in education lists the questions to ask.

Prompts, files and outputs in Magai are never used to train any model, from any provider on the platform. Keep the identifying details out anyway. A district rule doesn't care what a vendor does with training data.

Then read the whole week once as a student who has never seen the speech. Where would you get lost? That is where the gloss, the sentence frame or the extra two minutes goes, and it is a judgment only you can make.

If you want a wider look at what else teachers are using, AI-powered tools for educators covers the classroom side. This post is about the Sunday side.

Save the setup so next Sunday starts at the skeleton

You typed everything above once. Next week you shouldn't type it again.

Put the facts that don't change into an agent. An agent in Magai is a set of reusable instructions that runs on any model, so the same setup follows you whether you draft with /Claude and audit with /GPT. How to build an AI agent walks through making one, and these instructions are a starting point:

You plan ELA lessons for my 8th grade classes. Periods are 50 minutes with 28 students. About six read well below grade level and two are English learners. Always work from the reading I attach and quote it exactly, never from memory. List the minutes for every block and make them add up to the period. Every exit ticket must be answerable only by someone who read the text and must measure the standard, not recall. If a reply is running long, stop at the end of a day and tell me what is left.

Every line in that block answers something that broke in the run. "Quote it exactly" is there because the quotes are what you can't see being wrong. "Make them add up" is Friday's summary line. The exit ticket rule is the six look up questions. And "stop at the end of a day" is the cut off, handled before it happens.

Set it up in this order:

1. Create a new agent and paste the instructions, with your own grade, period length and groups.
2. Open a new thread inside it and attach the reading for the unit.
3. Paste the standard and send the first message from earlier in this post.
4. Create a folder named for the unit and move the thread into it. Folders organize your chats, so week two sits beside week one with the reading, the fixes and the audit already in it.

The next time the week comes round, the whole job is one message:

Here is this week's reading [attach] and standard [paste]. Same class and constraints as before. Give me the one line per day skeleton first, and don't write the lessons yet.

If you teach, Magai for education has more on how teachers set this up.

A good sub plan names the room. Yours should too. Open a new thread at Magai, attach the reading, and paste the standard.

Start there.