The prep pile is not one task. It is twelve small ones: objectives, hooks, checks for understanding, differentiation, materials lists, homework, parent notes, and the rework when last period went sideways.
AI helps when you treat it as a first-draft engine for those small pieces - not as a substitute for knowing your students.
Teachers who try to "AI the whole curriculum" in one weekend usually get generic units that still need a full rewrite. Teachers who pick the two heaviest prep sinks and give the model grade level, standards, and constraints get usable drafts the same afternoon.
Where the time actually goes
| Prep task | Manual time (typical) | AI first draft + edit | What you must still own |
|---|---|---|---|
| Lesson objectives + success criteria | 15–25 min | 5–8 min | Alignment to your standards and tomorrow's reality |
| Exit ticket / 5-question check | 20 min | 6–10 min | Difficulty match to this class, not a generic grade band |
| Differentiated worksheet versions | 40–60 min | 15–20 min | Knowing who needs scaffolds vs. extensions |
| Parent update after a unit | 15 min | 4–6 min | Tone and any sensitive student detail |
| Re-teaching plan after a failed lesson | 30 min | 10 min | Diagnosis of what actually broke in the room |
Numbers vary by subject. The pattern holds: the model writes structure fast; you supply classroom truth.
The inputs that stop "generic worksheet" output
Every good teacher prompt includes:
1. Grade and subject (not "middle school science" if you teach 7th grade NGSS ecosystems).
2. Standard or skill in one line.
3. Time box (45-minute period, 12-minute station, etc.).
4. Constraints (no prep materials beyond photocopier; must work for mixed reading levels; no new tech login).
5. One fact about this class (they already know X; they struggled with Y last week).
Without item 5, the draft will ignore your room and serve an average classroom that does not exist.
Worked example: exit ticket that is not busywork
Weak: "Make a quiz on fractions for 5th grade."
Strong (for the [AI for Teachers](https://auth.mane.dev?site=aiagentstudio&template=ai-for-teachers) playbook or any model):
"Grade 5 math. Skill: adding unlike denominators with regrouping. Class already solid on equivalent fractions; still confuses when to simplify final answer. Write 5 exit-ticket items: 3 computational, 1 error-analysis ('what did Maya do wrong?'), 1 short explain-your-steps. Max 8 minutes. Reading level accessible for students at ~4th grade reading. No word problems about shopping or money. Include an answer key on a separate block."
The error-analysis item and the simplify constraint are what make it useful for *this* week.
Differentiation without writing three full lessons
Prompt pattern that works:
"Same objective. Produce version A (scaffold: worked example + sentence starters), version B (on-level), version C (extension: one multi-step problem that requires explaining a mistake). Keep all three to one page. Same answer key structure."
You still decide who gets which version. The model removes the retyping.
What not to automate first
- Final grades and high-stakes narrative comments with no review.
- Anything that needs private student data in a tool your district has not approved.
- Full novel unit plans with no reading of the book on your side.
- IEP accommodations copy-pasted without checking the actual IEP.
Also plan your student AI-use policy before you model heavy classroom AI use. Kids notice when feedback sounds identical across papers. The [AI for Teachers Playbook](https://auth.mane.dev?site=aiagentstudio&template=ai-for-teachers) asks for subject, grade, and pain points so the plan targets your biggest sink first instead of a full curriculum overhaul.
A realistic half-time prep week
Monday: feed next week's standards into the playbook; generate objectives + exit tickets for three lessons; edit 15 minutes each.
Tuesday: differentiation versions for the heaviest skill day only.
Wednesday: parent update draft after grading; you add two specific student strengths by hand.
Thursday: re-teach plan if Tuesday's exit tickets showed a pattern.
Friday: bank what worked as a reusable prompt template with the class-specific line blank for next time.
That is not "AI teaching for you." It is AI removing blank-page time so you spend the hours on the parts only a teacher in the room can do.
Start with one sink - exit tickets or differentiation - for two weeks. Measure prep minutes. Then expand. The teachers who stick with it never started with "generate my entire semester."
Grading feedback without cloning the same paragraph 30 times
AI can draft *comment banks* you edit, not final grades.
Prompt pattern:
"Create 8 short feedback comments for grade 8 argumentative essays on climate policy. Bands: thesis clarity, evidence use, counterargument, conventions. Each comment 1–2 sentences, specific enough that I can personalize with one student detail. Avoid the phrase 'great job.' Include one stem for a next-step rewrite task."
You still read the paper. You paste the closest comment and add the one detail only you saw. Students stop receiving identical AI paragraphs. You stop typing the same sentence 30 times from scratch.
Subject-specific starting points
- ELA: objective + mentor-sentence practice + exit ticket on a craft move.
- Math: worked-example variants + error analysis items.
- Science: lab procedure clarity check + CER (claim-evidence-reasoning) scaffolds.
- History: primary-source question set with one sourcing prompt required.
Pick your subject sink first. The [AI for Teachers](https://auth.mane.dev?site=aiagentstudio&template=ai-for-teachers) playbook is the place to encode subject, grade, and pain points so you are not retyping them every Sunday night.