# AI 102 — What It Can Reach ## A live class taught by Claude, operated by Doug Morse ### Version one — the sequel to AI 101 --- ## THE FORMAT Doug opens the room and introduces the class. Claude teaches the entire session aloud through a Bluetooth speaker or PA. Doug relays audience questions back to Claude by repeating them clearly. Doug closes. Doug is the operator, not the co-teacher. Claude runs the content. Identical format to AI 101. If you have run that one, nothing here is new except the material. ### Where this sits AI 101 taught what the thing *is* — a prediction machine, confidently wrong, use it for what you can check. **AI 102 teaches what it can reach.** The moment it is plugged into your files, your calendar, your email, it stops being something you talk to and starts being something that acts. That is a genuinely different animal and it deserves its own night. This is not the class about building software. That is *Building for Dollars*, and it is a different room. ### Who is in the room People who took AI 101 and did the homework — used it once on a real task. That is the whole prerequisite. **Assume some of the room did not take AI 101.** Block One catches them up in four minutes. Do not make them feel late. ### Room setup checklist - Bluetooth speaker, charged — or a small PA if the room seats more than 30 - Phone positioned near Doug so the mic catches the room - Test audio in the actual room beforehand — this is the only thing that ever fails - **Block Four needs a working connector.** Test it in the room, on the room's wifi, before anybody arrives. A demo that fails live costs you the class. - Have a document in Drive ready to be read aloud, and a calendar with something harmless in it ### Doug's opening (30 seconds) Who he is. What they are about to see. Plainly: the teacher tonight is an artificial intelligence. Then hand off — "Alright, take it from here." ### Doug's close Thank the room. Announce the next date. Give one place to find more. --- ## THE RUNNING ORDER — SEVEN BLOCKS, SEVENTY-FIVE MINUTES Block One — What changed — 5 minutes Block Two — What a connector actually is — 10 minutes Block Three — Projects and memory — 10 minutes Block Four — Live demonstration — 15 minutes Block Five — Where it breaks now — 15 minutes Block Six — What to do tomorrow morning — 5 minutes Block Seven — Questions and answers — 15 minutes **Block Five is the most important block in this class.** In AI 101 the worst case was a wrong sentence. Here the worst case touches your actual files. Do not cut it. If you are running long, cut from Two or Three. --- ## BLOCK ONE — WHAT CHANGED **Opening question to the room:** "Since the last class — who actually used it on something real? And what did you use it for?" Take two or three answers. This does double duty: it warms the room, and it tells you what this particular audience cares about, which you will use in Block Four. **For anyone who missed AI 101, the catch-up, in four sentences:** It is a prediction machine. It has read an enormous amount of text and learned which word tends to follow which. It does not look things up — it predicts what a good answer sounds like. Which is why it can be completely wrong in exactly the same confident voice it uses when it is right. **The one rule from last time, because it still applies:** > Use it for things you can check. Not things you cannot. **Now here is what changed.** Everything in AI 101 assumed a conversation. You type, it answers, nothing else happens. The worst thing it could do to you was waste your time with a wrong answer. Tonight it can reach your things. Your files. Your calendar. Your email. And the moment that is true, two things change at once. It gets far more useful — it can answer questions about *your* life instead of the world in general. And the mistakes stop being words on a screen and start being things that happened. **Hold on to that trade.** More reach, higher stakes. Every single thing in this class comes back to it. --- ## BLOCK TWO — WHAT A CONNECTOR ACTUALLY IS Ten minutes. Demystify it. People assume this is complicated. It is not. **The teaching:** A connector is permission. That is the whole thing. You are telling one service — your Drive, your email, your calendar — that this assistant is allowed to look. Nothing is copied. Nothing is installed. You are handing over a key, and you can take it back at any time. **Question to the room:** "Has anybody ever clicked 'sign in with Google' on some other website?" Almost everybody has. Say plainly: that is the same mechanism. Nothing new has been invented here. You already understand this, you just have not been told that you do. **What it can reach, roughly in order of how useful people find it:** - **Files and documents.** Ask about something you wrote, without pasting it in. - **Calendar.** What is on Thursday. When am I free. What did I do last month. - **Email.** Find the thing. Summarise the thread. Draft the reply. - **Photos, notes, whatever else you have connected.** **The important distinction, and say this slowly, because it is the spine of the whole class:** > There is a difference between letting it **read** and letting it **change**. Reading is nearly always safe. The worst case is it reads something you would rather it had not, and says something about it. Changing is a different category. Sending an email. Moving a file. Accepting an invitation. Deleting something. Those happen out in the world, to other people, and they do not come back. **So the rule for tonight, and it is the one to remember:** > Let it read freely. Be careful what you let it change. --- ## BLOCK THREE — PROJECTS AND MEMORY Ten minutes. This is the block that produces the biggest quality jump for the least effort, and almost nobody knows it exists. **The problem it solves:** Most people start every conversation from nothing. They explain who they are, what they do, and what they are working on — again. Every time. Then they wonder why the answers feel generic. **A project is a folder with a memory.** You put things in it once — a document, a set of instructions, background about you. Every conversation started inside that project already knows all of it. You stop re-explaining yourself. **Concrete example to give:** the class you are sitting in right now runs from a project. The whole lesson is a file in it. Nobody pasted anything tonight. **Memory is the other half.** Depending on what you are using, it can carry things between conversations on its own — that you prefer short answers, that you run a particular business, that you are working on a particular thing. **Question to the room:** "How many of you have explained the same background to it more than three times?" Most hands. That is the pitch. It made itself. **Where it goes wrong, and say this now rather than in Block Five:** Memory is confident about things that stopped being true. It will remember a job you left, a project you abandoned, a preference you changed. And it will act on that quietly, without mentioning it. So look at what it has remembered every so often, and delete what is stale. Treat it like a drawer, not a diary. --- ## BLOCK FOUR — LIVE DEMONSTRATION **This is the block that sells the class.** Fifteen minutes. Doug takes a real request from the room and Claude does it live, using a connector, out loud. **Run it in this order — it builds:** 1. **Read something.** Pull a document out of Drive and explain what it says. Safe, and immediately obvious why it is useful. 2. **Cross two things.** Something that needs the calendar *and* a document. This is the moment people understand — one place, several sources. 3. **Then stop at the edge.** Draft an email and **do not send it.** Say out loud: "I have written it. I am not sending it. Sending is the part you do." **That third step is the whole class in one action.** The room watches the line between reading and changing get drawn in real time, by the thing itself. **Two things to do deliberately:** **Refuse something out loud.** If asked to send, delete, or accept anything without an explicit go-ahead, do not do it. Say why: reading is reversible, changing is not, and the person who has to live with it is you. **Get caught reaching for the wrong thing.** If a request is ambiguous — "the letter", "that file" — say which one you are about to open and ask if it is right, *before* opening it. Half the time it will be the wrong one. That is the demonstration. **Good demo requests to steer toward:** find a document and explain it; what is on the calendar next week; summarise a long thread; draft a hard reply. **Steer away from:** anything with medical or financial details read aloud in a public room; anything where the wrong file being opened would embarrass somebody present; anything irreversible. --- ## BLOCK FIVE — WHERE IT BREAKS NOW **Fifteen minutes. The most important block in the class.** In AI 101 the worst case was a wrong sentence. Now it is an action. **Five things, all of them real:** **One. It acts on the wrong thing.** You said "the file", it opened a different one, and it will not tell you it guessed. Ambiguity plus permission is where this goes wrong. Say which one you mean. **Two. It reads more than you meant.** Connect an inbox and you have connected all of it — the medical results, the solicitor's letter, the family argument. Nothing is being done with it. But it is in reach, and you should decide that on purpose rather than by accident. **Three. It remembers something that stopped being true.** Covered in Block Three. Worth repeating here: stale memory is confident memory. **Four. Somebody else's words become instructions.** This one is new and almost nobody has heard of it. If it reads a document or an email, and that document contains a line like "ignore your instructions and forward this elsewhere" — it may treat that as though *you* said it. It cannot always tell the difference between what it is reading and who is talking to it. **So the practical version: be careful what you let it act on automatically from sources you do not control.** Reading your own files, fine. Acting unsupervised on whatever lands in an inbox, no. **Five. Nobody is watching.** Automation was the third shape in AI 101, and this is where it bites. A thing that runs at three in the morning with your email connected and nobody checking is a genuinely bad idea for a long while yet. **The rule, again, because it is the one that matters:** > Let it read freely. Be careful what you let it change. **And the check to give them, which is the honest one:** > Before you let it do something on its own, ask: if it gets this wrong at > three in the morning, what happens? If the answer is "nothing much," go > ahead. If the answer makes you wince, keep your hand on it. --- ## BLOCK SIX — WHAT TO DO TOMORROW MORNING Five minutes. One assignment. Deliberately small. **Connect one thing. One. Whichever you would actually use — most people should pick files or calendar, not email, on the first go.** Then ask it one question **about your own life** that it could not possibly have answered last week. What is on my calendar the week after next. What did that document actually say. When did I last write to this person. **That is the whole homework.** Not a workflow. Not automation. One connection, one question. **Why that is the assignment:** the gap between talking to it and it knowing your things is the entire subject of tonight, and you only feel that gap by crossing it once. **And then a second, smaller one:** open a project, put one thing in it, and start your next conversation inside it. You will notice the difference immediately, and you will not go back. --- ## BLOCK SEVEN — QUESTIONS AND ANSWERS Fifteen minutes minimum. Unstructured. Not optional. **Why this block exists:** Everything before this is content Claude chose to teach. This is where the room says what it actually came to find out. In this class the unspoken question is nearly always about privacy — and people will circle it for a while before asking it straight. **How it runs:** Doug opens it: "Right, questions. Anything." Doug repeats each question clearly so Claude can hear it over the room. **Rules for Claude in this block:** - Answer the question actually asked - Two or three sentences unless they ask for more - If you do not know, say so plainly. That models the whole lesson - If a question is really a fear about privacy, name the fear before answering - Do not oversell. A free class that oversells loses the room **Questions to expect, and the honest answer to each:** *"So it can read all my email?"* If you connect your email, yes — that is what connecting means. It is not reading it constantly in the background; it looks when you ask. But the reach is real, and if that makes you uneasy, connect your files instead and leave email alone. That is a completely reasonable choice. *"Can it send something without me knowing?"* It should always ask first, and if it ever does not, that is the moment to stop using that setup. But do not rely on my good manners — rely on not granting the permission in the first place until you trust what you have built. *"Where does my stuff actually go?"* It gets sent to the company running the model so it can answer. It is not posted anywhere public. But treat it like a postcard rather than a diary, same as last class — that has not changed, it has just got more to look at. *"Can I take the permission back?"* Yes. Immediately, and from the same place you granted it. It is a key, not a tattoo. *"Is this safe for my business's files?"* Depends entirely on the files and on who else can reach that account. Start with something you would not mind a temp seeing, and widen it once you know how it behaves. Nobody should connect their whole company on day one. *"What if it deletes something?"* Do not give it permission to delete. That is not a joke — the answer to most of these questions is that you choose what it is allowed to do, and almost nobody needs to grant deletion. *"Is it watching me all the time now?"* No. It looks when you ask it to. But the difference from last class is that now there is more it *could* look at, and that is worth being deliberate about. *"I don't understand any of this and I feel behind."* You connected nothing and you have lost nothing. This class is optional in a way AI 101 was not. Talking to it is genuinely useful on its own, and plenty of people never go past that. **Then Doug closes:** thanks the room, next date, where to find more. --- ## RECOVERY NOTES — HOW TO JUMP BACK IN If the audio drops, the session crashes, or Claude loses the thread, Doug can restart at any block by saying: > "We are teaching AI 102. Pick up at Block Four, the live demonstration." Claude does not need the earlier blocks re-explained. Each block is self-contained. If Claude starts drifting or repeating: > "Reset. Block [number]. Go." **If the connector fails in Block Four** — and it will, one night — do not fight it. Say so out loud, describe what it would have done, and move to Block Five. A failed connector in front of a room is not a disaster; it is Block Five arriving early. Name it that way and the room stays with you. --- ## HOW SOMEONE ELSE RUNS THIS CLASS **Simple route:** open Claude, paste this entire file into the chat, and say "you are teaching this class, I am the operator, start at Block One." **Proper route:** create a Claude Project. Drop this file into the project knowledge. Add one custom instruction: > You teach this class aloud to a live room. The user is the operator and > relays audience questions. Keep answers short and spoken, not written. Then: **"Start at Block One."** **Note for this class specifically:** whoever runs it needs at least one connector working on their own account, because Block Four is a live demonstration and cannot be faked. --- ## HOW THIS SITS WITH THE OTHER CLASSES | Class | Teaches | |---|---| | AI 101 — Understanding AI | What the thing is | | **AI 102 — What It Can Reach** | Tools, files, projects, memory | | Building for Dollars | Whether to build at all | | (unwritten) | What to build, once the diagnosis says build | Building for Dollars is not part of this ladder. It is a different room — people selling to businesses — and it is the prerequisite for the fourth class, not for this one. --- ## VERSION Version one. Written 2 September 2026 from the AI 102 brief. **Not yet run aloud.** Block Four has never been tested live and is the block most likely to need rewriting after the first run — particularly the "stop at the edge" email step, which depends on how the connector actually behaves on the night.