Real AI builds, tools and practices for educators, from a year of building classroom
materials with LLMs for a second-grade reading group and enrichment classes in San
Francisco. Presented at the Khan Lab School AI in Education Summit, October 24, 2026.
Everything on this page is free to copy, adapt and share.
Over one school year I built 133 things with LLMs for one classroom: word study for
learners with IEPs, customized reading passages, scaffolded comprehension tools,
goal-setting chatbots, project guides, progress notes, full units. Three of those
builds are worth your time, and every one of them needed a teacher to change
something before it worked with children. That change is the content of this page.
Nothing here needs a specific tool. Use whatever your school has approved, and if
your students cannot have accounts yet, the pattern still holds: the teacher prepares
and demonstrates, the students read, discuss, design, test and make.
The one-paragraph version, if you only take one thing
Paste this into the instructions of a project, Gem or notebook in whatever tool your
school has approved. Fill in the square brackets. The last two sentences are the
ones that matter.
You are helping me, a [grade or age band] teacher, prepare classroom materials. My class: [size, grouping]. This unit's learning goals: [goals]. Sound-spelling patterns taught so far: [list]. Formats I can use: [print, teacher display, read-aloud, hands-on]. Typical time per activity: [minutes].
Rules for every draft: use only the sources I attach or paste, and tell me when a request needs a source I have not given you. Label anything invented, such as a scenario, a game rule or an example, so it cannot be mistaken for a fact. Never write the learner's final explanation, answer or reflection. Never assign a reading level as if it were validated; describe the text features instead. When I describe a learner barrier, work from the barrier as I describe it, not from a diagnosis. End every draft with a short list of the claims, words and instructions I should check before I use it with children. If I paste a child's name or anything that looks like student records, stop and tell me, and do not use it.
This is the short on-ramp. The full
teaching thought partner setup
adds file checks, a style pass and a teacher check at the end of every resource; the
one-paragraph guide explains the
context card and the two lists of what goes in and what never does.
Three builds, and what had to change
Each case is told the same way: the classroom need, what AI helped produce, what the
teacher changed, what learners did, what transfers, and one honest limitation.
Case 1
Gold Rush
The classroom need
A San Francisco 1849 unit for seven- and eight-year-old readers around level M, where the biographies feed the play instead of sitting next to it.
What AI helped produce
An Oregon Trail-style Gold Rush game, built by talking to an LLM and tested with second graders
Group profiles of the 49ers, and biographies of historical figures including Mary Ellen Pleasant and Mifflin Wistar Gibbs
A bingo and scavenger hunt keyed to the biographies, and a Time Traveler field log for the stations
What the teacher changed
Every scavenger-hunt question checked against the biography text, so the answer was findable but not obvious
The generated biography pages debugged by hand before any child saw them
Fact, interpretation and game mechanic kept visibly separate, and the cast reviewed for who was missing
What learners did
Learners played one decision at a time, read the biographies to answer the hunt, and said what information they used to choose.
A place-based study of the first peoples of the San Francisco area and their relationships with water, for the same readers, inside a unit on how the city works.
What AI helped produce
Waterways of the First People, an interactive guide with knowledge packets and riddles, and a one-page water connections treasure hunt that sends learners back to it
A landmark design and reflection sheet with systems-thinking prompts, a Tinkercad design check, and a zine creator worksheet
A San Francisco city systems quiz with an answer key
What the teacher changed
The first version read like a game. The framing was rewritten around relationships with water, with specific community names and community-authored sources where they exist
Every claim checked for time period, place, attribution and present-day continuity
The field study at Crane Cove became observe, do not collect
What learners did
Learners authored the systems explanations, the landmark models in Tinkercad and the zines. The AI-assisted pieces were the organizer, the source packet and the design check.
A generated page about living communities needs a human check on every name and claim. The digital material only earns its place when it points learners back to local observation.
Learners reading around level L who still need to explain the relationship between a place and its water system, plus daily word work for learners with IEPs and a research project comparing myths with informational texts.
What AI helped produce
Customized reading passages and comprehension scaffolds, with the learning target held fixed
Practice sets for ch and tch, ge and dge, long and short vowels, and homophones
A research poster organizer, an independent poster guide and a model poster, plus a five-minute reading check-in
What the teacher changed
The science sources that came back too advanced for level L readers were replaced with a simpler set, each one checked against the readers it was for
Every practice word checked against the taught patterns; a model’s reading-level label treated as a guess, never as validation
Barriers described as observed; diagnoses kept out of the prompt entirely
What learners did
The reading project ended in art, a written response and a recorded reading. The word work ran daily, from checked sets.
Access to content and decoding practice are different design tasks. A prompt that blurs them produces a text that is neither.
The same five questions for every build
This is the process behind the case cards, and the fields on the build card. Run it on
anything you will reuse, share or hand to another adult. A one-off morning message
does not need it.
What was the classroom need?
A real task, a learner barrier or a missing resource.
What did AI help produce?
The original draft or build, the prompt, and the source it used.
What did the teacher change?
A visible correction, revision or design decision. This is where the value is.
What did learners actually do?
A de-identified work sample or a specific observation. Engagement, completion, understanding and independence are four different things; say which one you saw.
What can another teacher transfer?
A small template and a check for readiness. Plus one honest limitation.
Where the LLM cost more than it saved
Four places from my own record, and what I did instead.
Task
Status
What I did instead
Generated biography pages for the 1849 unit
Not responding
Debugged by hand; hunt questions tied to text I had checked
Science sources for the reading project
Too advanced
Replaced with a simpler set, each checked against level L readers
The first Waterways framing
Distorting the purpose
Reframed around relationships with water, with named communities and their sources
“Second-grade reading level” labels
Unverified
Treated as a guess; words checked against the taught patterns
Ready means: the key claims are checked, the task matches the learning
goal, the supports fit the barrier, and you can say what evidence you will look for.
One lesson then tests the resource. It does not prove a learning gain.
Keep the learning target fixed and change the route to it. The prompt asks for defined terms, evidence questions, an answer key with source locations, and a list of what to check.
The prompt behind the Gold Rush game and the CTWS organizers: one page, decisions kept with learners, invented rules labeled so they cannot be mistaken for facts.
A short practice text built from your phonics sequence, with every out-of-pattern word listed so you can check it. The model does not get to call it decodable.
Directions one step per line, sentence frames that point to the text, pre-teach cards, chunked passages, half-filled organizers and faded supports. Each one is short on purpose.
The exact sequence of seven prompts, and the iterative fixes, that built the Gold Rush (1849 San Francisco) classroom game. Swap in another era and reuse the engine.
Five review questions a colleague can run on your draft, a one-lesson trial log, and an exit slip. One lesson tests the resource; it does not prove a learning gain.
Two prompts that turn a Project or Gem into a planning partner for any subject and grade: a Quick Start you can paste in five minutes, and a Teaching Profile for when you have twenty.
The short on-ramp: one paragraph of instructions, one context card, and the two lists of what goes in and what never does. Works in Claude, ChatGPT, Gemini or whatever your school has approved.
Upload this to your AI Project and anything for colleagues, families or students gets a style pass: plain words, no hype, no invented observations, and a six-check final pass.
guideteachersedubuildswritingcommunication
The full prompt library
Fifty-four prompts in eight groups: project setup and reusable context, spoken ideas
and research, adaptation and classroom artifacts, bulk data, family newsletters,
admin follow-through, student-note documentation, and independent student projects.
Each one works from your own files and ends with a teacher check.
Anything students, families or colleagues will read gets a style pass first. The
tone and voice style guide for educators
is the file the setup prompts call for by name: plain words, no startup language, no
invented observations, and a six-check final pass. Fill in the Make it yours table, then
upload it to your Project.
If you teach K-12 in the United States, both
Claude for Teachers
and
ChatGPT for Teachers
are free individual sign-ups for verified educators. Your school does not need to
enroll. Check your district's guidance on which tools are approved for planning before
you build, and keep student data out of all of them (the setup guide covers how).
Also at the summit
This session left several things out on purpose, because other presenters at the
Khan Lab School AI in Education Summit
covered them better than 45 minutes could:
Kanishka Seth on step-by-step AI instructions on the assignments you already have.
Angela Chen on which model actually sits behind a classroom tool, and how to vet it.
Azade Sanjari on hint ladders and help that does not do the thinking.
Tiffany Wong on K-5 inclusive design with AI, grounded in UDL.
April Lehman, Anuja Agarwal and Natasha McKeown on NotebookLM as a source-grounded notebook for learning and school communication.
Coming to this page. The slide deck, once it is final. Everything else
from the session is already here.
Bring this to your school
I run this as a hands-on workshop for schools and teams: real builds, a twelve-minute
build block, peer review, and a trial plan everyone leaves with.
Get in touch. And if you have a build that survived your own
classroom, send it in. This library grows by what teachers
actually used.