Friday, October 9, 2026 · Google Meet · two hours
Hands On, the first run
The kit's first run, for MNGAIA · AI4MN, where so many people had already built something that the morning became builders showing each other what works.
Overview
The short version
Three people showed tools they already use at work, four breakout rooms reported back, and Dan closed with textbooks that AI builds for students who never touch an AI chat. The same moves came up in every room: just ask, let the AI interview you one question at a time, paste the error back, and go a step at a time. So did the same two cares: meet people in the tools they already use, and treat student data as part of every build. It ended on sharing: Sheets that carry their scripts, GitHub for work others can undo and build on, and offers of time from people who've done it.
- 22questions, from the room and researched afterward
- 6themes, in the room's own words
- 9ideas, and the offers that answer some
- 35things shown and shared
Were you there?
The index has everything you saw, with links, and offers and asks says who to ask for what.
Weren't there?
Start with the themes, then open the questions that catch your eye; each one stands on its own. Want the whole two hours? Gemini's notes for the meeting hold its summary and the full transcript.
How the morning went, for anyone hosting
The plan is on Hosts. Here's what happened instead, in order.
- Welcome ran on the chat, as planned, with the stage's wait time doing the counting: a minute for wishes, then the 1-to-5 (seventeen answers in the chat: four 1s, five 2s, a 2.5, three 3s, three 4s, and a 5+; Sean suspected the 2s were 3s), then three minutes for the most useful thing people had built, which showed a room full of builders. Dan's browser said "this site can't be reached," so Dave posted the starter code deck in the chat.
- The three levels were named, not shown. Sean skipped the live demo ("I'd rather see what some other people are doing"), and Show & Tell grew to three people: Kate, Ken, and Paul. Builders raised a hand in Meet, and a host made them co-hosts so they could share their screens.
- Sean posted the Breakout Planner in the chat, and everyone signed up for a level, so each room would sit in people's zone of proximal development: people who would help them learn, and people they could inspire. Amanda turned level 1 into a room for anyone who couldn't open the site, working from the starter code slides, and Dave made a room for people who hadn't signed up, to talk through how others built things. When the rooms opened, Dave posted the Three levels page in the chat.
- The rooms ran about half an hour. One person missed Meet's notice for their room while signing up for a GitHub account, with the notice hidden behind other tabs; moving latecomers took two hosts; and the room timer was set but not started until halfway through.
- As the rooms came back, Sean's camera broke, so Dave started the room reports while Sean rejoined. Then came four room reports, three more demos (Kate finished a demo copy of her lookup while the reports went on), and Dan's textbooks; Sean thanked him afterward "for jumping in, cold." Grant a wish (now Invent your own), Show & Cheer, and the wonder and worry walls didn't happen, and the walls deck is still empty.
- Amanda had invited a cultural family advocate from her district the day before, because he'd been talking about AI. He came.
- Sean, at the close: "It was different than what I imagined, and I think it was great, and it was fun."
For next time: open the shared doc for anyone to edit before the day (Sean noticed it wasn't, partway through Welcome); start the room timer when the rooms open (and see smoother breakout rooms, researched afterward); give a level 1 room to anyone the site is blocked for; have a co-host ready to run the room reports; and when the chat shows a room full of builders, give Show & Tell the time and let the plan bend. Dave and Sean agreed that breakout rooms are not Meet's strongest feature.
Questions, answered
Getting unstuck
The AI gave me an error, or something I don't understand. Now what?
Room 1, reported by Amanda; Sean, after room 3 · In the recording at 1:13:41
Tell it exactly that: "I don't know what to do with this. Break it down for me step by step." Paste the error message itself into the chat, ask what it means, and ask it to fix it. Sean said that's how many of them learned to code: an idea, an error, the error back into the chat, a fix, with a lot of copy and paste in between. When an explanation flies over your head, ask for it in smaller chunks. You learn the code and how to learn with AI at the same time.
More: three moves for when you're stuck, on Grant a wish.
How do I write the right prompt?
Room 2, reported by Patrick; room 4, reported by Meghan · In the recording at 1:16:16
You don't have to write one big, perfect prompt. Ask the AI to ask you questions first ("What will you need from me to make this successful?"), answer them, and keep adjusting until it makes the thing you want; the more you do it, the more you know what to say up front. Sean's version, which he gave room 1: "I've got an idea for a code project. Ask me one question at a time to help generate the code." Room 4 called it just asking: say what you're looking for and what you want to do, and ask for help.
It worked, I added one more feature, and now it's broken.
Room 3, reported by Ken · In the recording at 1:18:58
Most builders know this one. Room 3 built the Drive inventory script from the starters, started over from scratch when it wouldn't run the way they wanted, got it working, and then broke it adding a feature. Ken's habit: ask for a step-by-step plan, do the first step, attach the file back into the chat, and ask "check my work: does this look right?" so the chat keeps you on track. Sean's: the hardest part is taking your hands off the code instead of adding one more thing. Dan's: version control, so you can always go back (below).
More: the Drive inventory script, on Three levels.
Building and sharing
How do I put an interactive page on a Google Site without the frame around it?
Sean · In the recording at 14:02
Embed the HTML as a page of its own instead of as a box on a page. Sean mentioned it in passing at an earlier event, and a lot of people told him it was the useful part, because it got rid of a header they couldn't otherwise remove. Google Sites calls it a full page embed, added from the Pages tab; Google's announcement has the steps. A flat document becomes an interactive page on your site, quickly.
More: level 2, on Three levels.
How do I make a tool someone else can set up on their own?
Sean, after Ken's report portal · In the recording at 1:35:47
Ask the AI to build a setup wizard that walks the person through it one step at a time, so nobody has to find the line in the code where the folder ID goes. A sidebar in the Sheet, like the one Kate showed, gives people menus and buttons instead of code. Then ask: "Imagine I am experiencing this for the very first time. What steps would someone need to be successful using it?" and design those steps with it. Sean: those are the very sentences he uses; it's plain language, saying what the need is, until the tool can do the work.
Where do Apps Scripts start to break?
Sean asked Paul · In the recording at 27:44
When the data gets heavy. Paul's peer-evaluator app, where observers worked through the Danielson rubric and saved long notes into a Sheet, kept working all year but got visibly slower: the Sheet can hold it all, but pulling it back out into the web app takes longer and longer. Keeping some of it in the browser's cache helped, not enough; Paul thinks that one has outgrown Apps Script. Sean's way around it, for a tool many people use: give each group its own copy, so no one Sheet carries everyone.
Same moment, later: In the recording at 1:27:41
Researched afterward: the documented limits.
Student data
Does student information go into the Sheet behind a script, and who gets to see it?
Patrick asked Paul, in the chat, during Paul's demos
Only the pieces the tool needs, and only for the people whose role calls for them. Paul keeps those pieces (a student's name, a guardian's name and email) in a separate protected sheet, and the web app shows each person what their role allows: teachers see only their own students, and administrators a wider group. Patrick's follow-up, whether a hidden, protected sheet with a column on each row decides what the web app shows, didn't get an answer in the chat. It's a good one to ask Paul.
More: student data, on Sources.
Researched afterward: is a hidden, protected sheet safe?
How do I give families, staff, and the business office the privacy information each one needs?
Room 4, reported by Meghan; Sean · In the recording at 1:23:09
Room 4's advice was to just ask: describe the three audiences and what each needs, and ask the AI to help write the script. Sean showed where asking leads. His privacy inventory began in AI Studio with one question: can you take a privacy policy's web address and make it readable? Then one more ask at a time: measure it against federal and state law, list the data each tool uses, publish a library families can browse, add a request portal for staff, and generate a data privacy agreement to download and send the vendor. Eighteen months on, he's learning to open-source it so others can start from it.
More: student data, on Sources.
Researched afterward: what Minnesota's law asks districts to tell families.
How do we get score reports to families securely when the student information system can't?
Ken · In the recording at 1:31:48
As his district's assessment coordinator, Ken had to send families the actual MCA reports, one PDF per subject per student, in a way that keeps to FERPA. Infinite Campus can attach them only one at a time, an add-on didn't make sense for a district his size, and mailing them takes too much time. So he built a portal in about two days with Gemini, as a script bound to a Sheet. The PDFs go in one Drive folder (each file name carries the student's MARSS number, as MCA sends them, and a built-in renamer handles the rest); guardians and their emails go in the Sheet; and the setup steps walk you through deploying it as a web app. Families get the link by email from the student information system, enter their email, receive a one-time six-digit code, and download only their own student's files. Adding a language code to the link shows the page in Spanish or Somali. He has used it for MCA and ACCESS (WIDA) reports.
More: the report portal, which Ken offered to share.
Researched afterward: what FERPA asks of electronic delivery.
Going further
Can students learn from AI-made material without using AI themselves?
Dan · In the recording at 1:42:02
That's how Dan builds intelligent textbooks: AI writes the book, and students never use a language model. The pages are interactive on purpose. A picture is never just a picture: students click to hear a story or explore a part, and Dan used one AI tool to make the image and another to add the clickable layer on top. Every click is a chance to see whether a concept has landed. Dan calls these level 2, the way self-driving cars have levels; fully autonomous is years away.
More: Dan's textbooks, in the index.
How would a book know what a student understands, without a quiz?
Dan · In the recording at 1:51:52
Each book has a brain: a graph where every concept is a point, and the order you'd teach them is the line between two points. What a student does on the page suggests which concepts they've mastered, so a teacher can step in during week two instead of finding out at the end of the course. Dan's two cautions: that activity is student data, so where it's stored matters (Canvas keeps enrollments and exam scores, but doesn't model concepts); and none of this has yet been tested against a paper textbook, which is why he's looking for a university to study it.
Researched afterward: measuring from activity, and what Canvas does track.
What comes after Apps Script?
Dan; Sean · In the recording at 1:48:19
Version control, first. Dan's rule: before anything else, set up the teacher's computer so everything is undoable, with GitHub. Then the AI can build boldly, you can say "I don't like everything that you did since last Tuesday" and it's undone, and students get stable versions. After that, the desktop: an AI agent (Dan named Claude, ChatGPT, Google Antigravity, Devin, and AI running on your own hardware), an editor to see what it makes (VS Code), and on Windows, the Windows Subsystem for Linux. Plan on about an hour for Windows and twenty minutes for a Mac. It helps to know the idea underneath: a big problem broken into small ones until each has a recipe. Sean is writing a guide to GitHub for teachers; for him and Paul, version control was the breakthrough.
More: Dan's desktop setup (desktop tool, GitHub, VS Code, Unix shell), and level 4, the workbench, on Three levels.
Researched afterward: version control for Apps Script.
Researched after the session
Questions the room raised or brushed past, answered afterward from primary sources: Google's documentation, the U.S. Department of Education, Minnesota's statutes, and published research. These are our reading of those sources, not what the room said, and each lists what it rests on.
Could an event check-in app help reunify families after an emergency evacuation?Researched
Nathaniel, in the chat, after Kate's and Paul's check-in tools. Not answered on the day.
It could help with one piece: a running record of which students are where, and when each one was released. It shouldn't be what decides who a child goes home with. In the Standard Reunification Method, which many districts use, students wait apart from arriving parents; each parent fills out a reunification card, staff check photo ID and custody, a runner brings the student, and a final step accounts for every child. Federal guidance says a plan should verify that an adult is authorized to take custody. The method runs on paper, and reunification can follow a power outage, so a plan that uses an app still needs a paper way it has practiced. Minnesota requires every school board to adopt a crisis management policy, written with law enforcement and emergency managers; the statute lists drills but doesn't mention reunification. Your district's crisis team decides what a tool may do, not the person who built it.
What limits do Apps Script and Google Sheets actually have?Researched
Behind Sean's question to Paul about where scripts break.
Each run of a script gets six minutes; a custom function used in a cell gets thirty seconds. Daily allowances are per person, and a Google Workspace account (school accounts are Workspace accounts) gets more than a personal one: 1,500 email recipients a day instead of 100, and six hours of trigger time a day instead of ninety minutes. One person can run thirty scripts at once. Since September 2026 a spreadsheet can hold 20 million cells, up from 10 million, but a Sheet slows down long before it's full. Google's advice is to read a whole range in one call, work on it, and write it back in one call, rather than cell by cell. Apps Script's own cache holds up to 100KB per item for at most six hours, and may let go sooner. Google says quotas can change, so check its page before planning around a number.
What comes along when someone copies a Sheet that has a script?Researched
Behind Sean's "many hands make light work," and Ken's portal, which travels this way.
The script does. Whoever makes the copy owns it and gets their own copy of the script, and their changes never touch yours. What's tied to a person stays behind. Installable triggers, such as a timed run or a run when a form is submitted, belong to whoever made them, so the new owner sets them up again. A web app is a deployment, so the copy has none until its owner deploys it, and then it has its own address. The first time the new owner runs it, Google asks them to approve what it can reach. A share link ending in /copy instead of /edit opens a Make a copy prompt. Before sharing, clear out the data and remove any keys or settings stored with the script. A district can turn Apps Script off or limit sharing outside the district, so a copy that runs for you may not run for a teacher in another district.
Can an Apps Script project use version control, like GitHub?Researched
Behind Dan's advice to set up version control first, and Sean's guide to GitHub for teachers.
Yes, two ways. In the Apps Script editor, Project History keeps the versions you create, usually one each time you deploy, up to 200, and can highlight what changed. It isn't a record of every save, and going back means copying the old code in by hand. For full version control, Google's free tool clasp copies a project's files to your computer, where they can live in a Git repository on GitHub. clasp pull brings the online code down; clasp push sends yours up and replaces what's there, so edit in one place at a time. It needs Node.js 20 or later, which usually takes administrator rights to install, and the Apps Script API turned on in your Apps Script settings. On a district laptop that may be a request to your technology office: it's level 4, the workbench.
More: level 4, the workbench, on Three levels.
What this rests on, checked October 12, 2026:
What does FERPA ask of a school that sends score reports to families electronically?Researched
Behind Ken's secure report portal.
Parents have a right to see their child's education records. Before a school discloses records by any route, FERPA asks it to use reasonable methods to identify and authenticate the person receiving them. The rule doesn't name a method. Federal guidance asks for certainty that fits how sensitive the data is, and describes the factors: something you know, something you have, something you are. A one-time code sent to an email address is one such check; how much it proves depends on where the address came from, who else can read that inbox, and whether a family can reach only its own student's reports. Either parent keeps full rights unless a court order or similar legal document says otherwise. In Minnesota, the state treats these reports as private student data, and districts keep them secure until they reach families. The district chooses its method and answers for it; this isn't a ruling on any one tool.
What does Minnesota's law ask districts to tell families about technology providers?Researched
Behind Meghan's three audiences and Sean's privacy inventory.
Since the 2022-23 school year, Minnesota's educational data law, section 13.32, has rules for technology providers that handle student data under a school contract. That data isn't the provider's property. The provider can't sell or share it or use it commercially, must tell the school about a breach, and must return or destroy the data within 90 days after the contract ends unless renewal is expected. Within 30 days of the start of each school year, a district must give families direct notice, by mail, email, or something like them, of each curriculum, testing, or assessment technology provider with access to student data, which data, and whom to ask; families can inspect the contracts. A separate part limits remote monitoring of school-issued devices (location, camera, microphone, keystrokes, browsing) to listed exceptions. That covers the families; staff and the business office are the district's to decide, with its data practices official.
More: 13.32, on Sources.
What this rests on, checked October 12, 2026:
Can learning be measured from what students do, instead of from quizzes?Researched
Behind Dan's intelligent textbooks, and ahead of the November session on assessment.
Partly, and researchers have worked on it for decades. Since Corbett and Anderson's 1994 paper, intelligent tutors have used knowledge tracing: each step a student takes updates the tutor's estimate of whether they know a skill. Valerie Shute's stealth assessment builds the measuring into a game or task, so students show what they know by doing it. Both rest on careful design: a map of the skills, much like a concept graph; a plan for which actions count as evidence of each one; and studies showing the estimates agree with other trusted measures. Clicks count only when the task actually requires the skill, and these models can carry bias. One correction to the record: Canvas does track mastery, through Outcomes and the Learning Mastery Gradebook, scored from rubrics and assessments a teacher aligns to standards. What it doesn't do is infer mastery from page activity, which is the gap Dan described.
More: assessment, on Sources.
How can a host make Google Meet breakout rooms run more smoothly?Researched
Behind the morning's breakout rooms: a missed notice, latecomers to move, and two hosts agreeing they're not Meet's strongest feature.
Do most of the setup before the meeting. Check your edition first: Google's help lists Education Plus and the Teaching and Learning Upgrade among those that can create rooms, so not every school account can. A host can build the rooms in the Google Calendar event and pre-assign people by typing names, dragging, or shuffling. Turn on Host Management and add co-hosts beforehand: its settings can't change once rooms are live, and people can be made co-hosts only from the main room. Each participant gets a prompt on screen and must click Join; anyone who cancels stays in the main room, so say out loud and in the chat that rooms are opening, and ask everyone to bring the Meet tab to the front. Set the timer, which every room can see. Hosts can join any room, and a room can ask for help.
More: Hosts, for running the session with your own group.
What this rests on, checked October 12, 2026:
Themes
What kept coming up, across rooms that never met. The quotes are word for word from the transcript or the chat, with ums and repeats taken out, and credited to the moment rather than the person.
Just ask, then let it interview you
Every room landed here on its own. Say what you need, let the AI ask you questions one at a time, paste the error back, and go a step at a time; ask for smaller chunks when an answer flies past you. The limit is often one we put on ourselves, and the moves carry beyond code: they're how anyone learns with AI.
We all felt kind of like newbie beginners. So we just embraced our newbie beginnerness.
We've seen wizards do it, right? Like people that we admire who are able to code solutions, and we hold them on a pedestal.
Meet people where they already work
The tools that stuck live inside something people already open: a Sheet teachers already use, an email from a system families already know, a page in the format they need, whether that's read-aloud, Spanish or Somali, or a dictionary without the ads. One good experience in a familiar place opens the door to the next, and the surest sign a tool works is that nobody mentions it anymore.
How can we give teachers some efficiency tools that exist without revamping their entire world in a day?
I don't hear much about that one anymore because it's pretty smooth sailing and people tend to like it.
Start small, and know the edge
Each level seems impossible until it works, and then the next one does too. A walkthrough app took twenty minutes; a year of observation notes slowed a Sheet until it dragged. Know where a script stops, keep versions you can undo, and expect the ceiling to keep moving, remembering that all of it is new and still untested.
When you experience the first one, you're like, this won't work. And you're like, oh my gosh, that works. Then you try the next one and you're like, oh, that works too.
Just when you think you find the limits… the models get better… and you also learn a lot in that process.
Student data is part of every build
Privacy wasn't on the agenda, and it came up everywhere. Demos used made-up students, a portal sends one-time codes, a protected sheet shows each role only its part, an inventory turns vendors' policies into plain language, and a textbook's mastery data needs a home. Nobody on the planning committee is a software developer, so learning to do this safely is part of the work.
We use AI to generate the textbook. We don't have the students use LLMs. Very, very key difference.
It didn't say I could print it. It said I could download it, because it's not going to tell me to print stuff.
Share it, so nobody builds it twice
Two districts built ways to check students in at events without knowing about each other. A Sheet that carries its script, a GitHub repository, and more than a hundred open textbooks are all ways to take an idea, fit it to your place, extend it, and share it back. The chat was already asking where check-in could go next: reunification after an evacuation.
Grandma Mary used to say many hands make light work.
If you say, "I don't like everything that you did since last Tuesday," poof, it just undoes it.
Hard to stop, so ask for joy
A 5 on the 1-to-5 meant "a difficult time stopping," and the chat agreed. Building can be absorbing enough to cost sleep, so two habits help: tell the AI the feeling you want (fun, joyful, satisfying), and know when to take your hands off the code.
There may need to be an intervention where somebody comes in and tells me it's time to take your hands off the code… and go to bed, which perhaps happened last night.
I only sleep when my tokens expire!
…you had way too much fun designing this interface. I love it.
Offers and asks
Who offered what, and what people asked for, out loud and in the chat.
Offered
- Dan: an hour on Google Meet with any Minnesota teacher who wants to explore his books and tools, and mentoring for anyone who wants to make their own textbook: a six-week process, an hour a week. In the recording at 1:39:54 When Meghan asked how to reach him, he shared his LinkedIn.
- Ken: a copy of his report portal, as a Sheet with the script inside. He's updating its template from a wide data model to a tall one. In the recording at 1:33:49
- Ken: links to his examples, including Intervention Manager for PBIS and ClassPass. In the recording at 1:30:56
- Paul: his Apps Script work, on his GitHub, and his Thought Leader Summit talk on Apps Script; Sean offered them for him. In the recording at 28:51 Links to come.
- Sean: his privacy inventory, open-sourced on GitHub once he's learned how, and a guide to GitHub for teachers, in progress.
- This site: fork it and run your own. In the recording at 3:09
Asked for
- Dan is looking for a college or university to run A/B studies of an intelligent textbook against a paper one. So far he hasn't found one. In the recording at 1:38:49
- Sean asked builders to share links to what they've made, so others can see it working. In the recording at 9:34
- Eric wants help refining a superintendent email agent that sorts what's for the superintendent, what's for the board, and what other administrators or principals can take on, with a morning brief.
- Room 4 wants a way to give families, staff, and the business office the privacy information each needs; it's Meghan's project for the week after. Where to start.
Ideas from the chat: nine
When Sean asked what people would build if they could. Invent your own is the place to start one. In the recording at 5:22
- Amanda: a mileage reimbursement calculated from her calendar.
- Troy: a project tool that tracks progress on goals.
- Erin: a Sheet where testing-center teachers enter the test and the students, which emails them when a test still isn't done after a week.
- Tia: "an expert grant writer at my fingertips that would be awarded every time."
- Dan: an influence graph for every school, college, and university in Minnesota, showing who is for AI and who fears it.
- Rachel: an analysis of time and productivity, with recommendations built in.
- Ken: printable hall passes for testing, made in a batch from the schedule.
- Elam: something that fills in account sign-ups, conference requests, grant applications, and web forms for you.
- Eric: a vetting tool that finds every stakeholder in a project and plans how to communicate and lead the change.
Coming up
- November: MNGAIA · AI4MN turns to assessment, with Donna, the group's assessment expert, building on a national conference in Bloomington, NCME's Classroom Assessment Conference (November 12 to 14, 2026), which Dave linked in the chat: back, as Sean put it, to "what we're actually working on in a day-to-day basis." In the recording at 1:56:21
- Scripts and dips: Sean's idea for meeting in person to build together. Not scheduled yet. In the recording at 1:55:14
- The recording is going to everyone who was there.