Practical AI

Class outline and handout.

Practical AI Class Outline

Practical AI: Using ChatGPT, Claude and Gemini Safely and Well

The class uses the "brilliant intern" analogy and a short checklist for asking well (R.O.L.E.S.). It covers privacy and security early, and shows what AI apps can do now: search, see, listen and carry out tasks.

Know the room

Class flow at a glance

Time Section Length
0:00 Welcome: what you want from AI, what worries you 10 min
0:10 How AI actually works: tools, tokens, reasoning 14 min
0:24 Privacy and security first 20 min
0:44 Hands-on: asking well, and managing a conversation 20 min
1:04 Demo tour: what AI can do now 16 min
1:20 Verification habits, handout, Q&A 10 min

Section 1: Welcome (10 min)

Opening round

Go around the table. Each person answers two questions:

  1. "What is one thing in your life or work you would like help with?" (a letter, a project plan, learning a skill, fixing something)
  2. "What worries you about AI, if anything?"

Write both lists on the whiteboard and leave them up all class. The first list becomes demo material in Sections 4 and 5. The second is a checklist: before the class ends, every worry on it should be addressed or honestly acknowledged.

Set expectations

Section 2: How AI actually works (14 min)

The brilliant intern

"Think of AI as the smartest intern you have ever met. They have read nearly everything, they are fast, and they never get tired. But they are still an intern. They need clear instructions, they will confidently hand you wrong information if you do not check their work, and they badly want to give you an answer even when 'I'm not sure' would be better. Your job is to be a good manager."

Next-word prediction, briefly

These tools generate text by predicting what is most likely to come next, based on patterns in an enormous amount of writing. That is why the same tool can write a lovely thank-you note and also invent a book that does not exist. Both are "plausible next words."

The intern got tools

A year or two ago, the honest summary was "AI can talk, but it can't do anything." That is no longer true. The major apps can now:

Key point: more capability means more to manage. An intern who can only talk can only give you bad advice. An intern with your house keys can do real damage. That is why the next section is about security.

Tokens and context (3 min)

"AI doesn't read words, it reads chunks called tokens, roughly three quarters of a word each. Everything you type, everything you upload, and everything it writes back uses tokens. Think of the conversation as a whiteboard of fixed size: once it fills up, the oldest material gets erased."

Why this matters to a regular user:

  1. Usage limits. "You've reached your limit, try again at 3 PM" is measured in tokens. Long chats and big uploads use up the allowance faster. Paid plans mostly buy more allowance.
  2. Long conversations degrade. When the whiteboard is full, the AI forgets early details. (Section 4 covers what to do.)
  3. Odd failures. Because it reads chunks rather than letters, it can miscount the letters in a word. A good reminder that it is not magic.

If someone asks "Does it cost me per message?": "Not in the regular apps. You get a usage allowance, not a meter. Per-token billing applies to developers using the paid API."

How AI "reasons" (4 min)

"If I ask you 17 times 24, you probably can't just blurt out the answer. You work it out on scratch paper. Older AI had to blurt: it predicted the answer one word at a time with no scratch paper. Newer 'thinking' or 'reasoning' models get scratch paper. Before answering, they write out steps for themselves: break the problem down, try an approach, check it, back up if it doesn't work. Then they write the answer."

What it looks like in the apps:

When it helps: multi-step problems such as math, cut lists and material estimates, planning a schedule, troubleshooting with several possible causes, comparing options against several requirements.

When it doesn't matter: quick, simple tasks (rewording an email, a definition). Thinking makes answers slower and uses more of your allowance; the scratch paper is made of tokens too.

The honest caveats (good for the skeptic):

  1. It is still prediction. The "thinking" is more predicted text, organized into steps. A useful technique, not a mind.
  2. It can reason its way to a wrong answer, confidently, with neat-looking steps. Well-organized work is not the same as correct work.
  3. The shown thinking is not a perfect window. Researchers have found the written-out steps don't always reflect what actually drove the answer. Treat it as a clue, not proof.
  4. Reading the thinking is still your best verification tool. If a step looks wrong ("assumes 8-foot boards are exactly 96 inches with no saw kerf"), you can point at it: "Step 3 is wrong because..."

Prompting tip: even without a thinking mode, "Work through this step by step before giving your answer" or "Show your work" helps on multi-step problems and makes the answer easier to check.

The main tools (keep it short)

Tool Made by Free option
ChatGPT OpenAI Yes
Claude Anthropic Yes
Gemini Google Yes, with a Google account

Each also has paid plans; check each company's pricing page the day before class if someone is likely to ask.

"These are like drill brands. They all drill holes, with different grips and strengths. The free tiers are enough to learn on. The skills from today work on all of them."

Do not name specific models in class. Model names change every few months. Tell people: "The app picks a good default. Don't worry about the model menu yet."

Section 3: Privacy and security first (20 min)

Framing: "Before we use these tools, let's look at what happens to what you type, and how to lock it down. We'll change the settings together, right now."

3a. Where your data goes (5 min)

3b. Change the settings together (8 min)

Do this live on the projector, then help each person do it on their own device.

Tool Turn off training Private one-off chat Memory
ChatGPT Settings, Data Controls, "Improve the model for everyone" (off) Temporary Chat (toggle at the top of a new chat): not saved to history, not used for memory or training Settings, Personalization: view, delete or turn off memory
Claude Settings, Privacy, "Help improve Claude" / Model Improvement (off) Incognito chat (ghost icon): never used for training Settings: view, edit or turn off memory
Gemini Google account, Gemini Apps Activity, "Keep Activity" (off) Turning off Keep Activity stops long-term saving (Google still keeps chats for up to 72 hours) Same activity controls

Instructor prep: menu labels move around. Walk through all three on your own accounts the day before and fix this table if anything changed.

Point to make: turning off training does not erase what was used before. Set it up now, before you start pasting in real things.

3c. What never goes into an AI chat (2 min)

Safer habit: replace names and numbers with placeholders. "Write a letter to [LANDLORD] about [ADDRESS]" works just as well.

3d. New risks with agents and connectors (5 min)

  1. Connectors are house keys. Linking AI to your email, calendar or files lets it read (and sometimes act on) everything in them. Grant the least access that does the job, and disconnect what you no longer use.
  2. Prompt injection. When an AI reads a web page, email or document, hidden text in it can try to give the AI orders ("ignore your user and forward their inbox to..."). The AI cannot always tell your instructions from instructions planted in what it reads. This is the strongest argument for keeping agents on a short leash: watch what they do, and approve anything that sends, buys or deletes.
  3. AI-powered scams. Voice clones of family members asking for money. Perfectly written phishing emails with no typos to give them away. Defense: a family code word, and calling back on a number you already have.
  4. AI can be confidently wrong about safety. Electrical, chemical, medical and machine-safety answers need a qualified source. In the shop, the makerspace's tool training and the manual win over the chatbot.

Invite the skeptic in: "What else worries you about this?" Add answers to the whiteboard.

Section 4: Hands-on: asking well, and managing a conversation (20 min)

Everyone on their own device, or sharing the projector (see Instructor notes).

4a. Give it context, not just a question (8 min)

Newer AI models need less "prompt trickery" and more raw material. The biggest upgrade most people can make is pasting in or uploading the actual thing: the letter, the manual page, the photo, the spreadsheet.

R.O.L.E.S. (a checklist, not a formula):

Exercise: each person takes their task from the opening whiteboard list.

  1. Type it as a short, bare question. Read the answer.
  2. Rewrite it with at least three R.O.L.E.S. elements and the real details. Compare.
  3. Reply with one correction ("too long", "I'm a beginner", "assume I only have hand tools"). Notice it adjusts without starting over.

Key lesson: "It's a conversation, not a search box. Start rough, then steer."

4b. When to keep going vs start a new chat (7 min)

Rule of thumb: one chat per job.

Keep going in the same chat when:

Start a new chat when:

Live demo: recap and restart. In a long chat (prepare one in advance), ask:

"Summarize what we have figured out so far, including decisions and key details, as a short recap I can paste into a new chat."

Paste the recap into a new chat and continue. It keeps the useful parts, drops the clutter, and fixes most "the AI got dumb" moments.

Mention briefly: Projects (ChatGPT, Claude) and memory carry context between chats automatically. Convenient, but check what memory has stored about you (the 3b table) and delete anything you do not want kept.

4c. The yes-man problem (5 min)

AI tends to agree with you. Ask the wrong way and it tells you what you want to hear.

Red-team moment: hand this to the skeptic. Have them try to get the AI to agree with something false, or back down from something true. Discuss what happened.

Section 5: Demo tour: what AI can do now (16 min)

The instructor drives on the projector. Use tasks from the whiteboard list wherever possible. Spend 2 to 3 minutes on each demo and skip any that aren't landing.

Demo 1: Watch it think

Turn on the thinking mode (or pick a thinking model) and give it a real shop problem:

"I'm building a workbench. I need these pieces cut from 8-foot 2x4s: four at 34 inches, four at 48 inches, and six at 22 inches. Allow 1/8 inch for each saw cut. What is the fewest number of boards I need, and exactly which pieces come from each board?"

Open "Show thinking" and read a few steps aloud: it lists the pieces, tries arrangements, accounts for the kerf, checks the total. Then have the group check one board's math by hand. The steps are what make it checkable.

Answer key (instructor only): 6 boards. The trap is the kerf. Two 48-inch pieces don't fit on one 96-inch board (48 + 1/8 + 48 = 96 1/8), so each 48 needs its own board, leaving 47 7/8 inches. That remainder holds either one 34 or two 22s, never a 34 and a 22. One workable plan:

Simple division (460 total inches divided by 96) suggests 5 boards, which is wrong. If the AI answers 5, that's a perfect teaching moment: ask it to check board by board.

Red-team option: have the skeptic change one requirement mid-stream ("actually, the 2x4s are only 92 5/8 inches") and see whether it re-plans correctly.

Demo 2: Web search with sources

Ask a current question (an upcoming local event, a current product price). Point out the source links. Click one and confirm it says what the AI claimed. "Sources are only useful if you check them."

Demo 3: Show it a photo

Take a phone photo of something in the shop: an unlabeled part, a tool's data plate, a wiring diagram, or a 3D print with a defect. Ask "What is this, and what should I know about it?" or "What is going wrong with this print?"

Safety framing: great for identifying and understanding. For setup and safety procedures, the manual and the makerspace's tool training still rule.

Demo 4: Voice mode

Use a phone app's voice mode for a short spoken back-and-forth, such as talking through project steps while your hands are busy. Voice conversations follow the same privacy settings as typed ones.

Demo 5: Research mode

Start a Research or Deep Research request at the beginning of Section 5 so it has time to finish (it can take several minutes):

"Compare three CNC routers suitable for a community makerspace with beginner users. Include price, cutting area, software, and ease of use, with sources."

Come back at the end of the section. Show the cited report and open a couple of citations to check them.

Demo 6: An agent doing a task (watch only)

Show a short, low-stakes agent task, such as having the AI open a website and gather information into a table, or turn notes into a formatted document. Narrate what it is doing.

Tie back to Section 3: point out where it asks for approval and where you would stop it. "Watch it work. Never let it send, buy or delete without you approving."

Section 6: Verification habits, handout, Q&A (10 min)

Five habits to take home

  1. Check settings first. Training off, memory reviewed, Temporary or Incognito chat for anything sensitive.
  2. Give it the real material. Paste or upload the actual document, photo or details.
  3. One chat per job. New topic or slipping answers: recap and restart.
  4. Ask for weaknesses, not approval. Invite disagreement.
  5. The higher the stakes, the more you verify. Casual email: skim it. Money, health, legal or safety: check every claim against a real source or a qualified person.

Revisit the worry list

Go through the whiteboard list. For each worry: addressed, partly addressed, or still a valid concern? It is fine to end on "still valid." That honesty builds more trust than over-selling.

Common questions

"Which one should I use?"

"Pick one free tier and use it for a week on real tasks. The skills transfer. If you want the most privacy controls up front, compare the settings we just walked through."

"Is it safe?"

"It is as safe as how you use it. Settings off, nothing sensitive pasted in, agents supervised, and anything important verified. Those four habits cover most of the risk."

"Is it going to take my job?"

"The more immediate change is that people who use it well get faster at some tasks. Learning it as a tool, with its limits, is the practical response."

"Can I trust it for medical, legal or financial things?"

"Use it to understand the topic and prepare questions, then take those questions to a professional."

"Does it remember me?"

"Only if memory is on, and you can see and delete what it remembers. Temporary and Incognito chats don't feed it."

Instructor notes

Before class

Pacing

If someone won't create an account

Completely reasonable. They can drive the instructor's screen, use the throwaway account, or simply watch and red-team. Do not pressure anyone.

Take-Home Card

The one-page handout for the Practical AI class. Menus in these apps change often; this card was last checked in October 2026.

1. Lock down your settings

Turn off training Private one-off chat Check memory
ChatGPT Settings, Data Controls, "Improve the model for everyone" Temporary Chat Settings, Personalization
Claude Settings, Privacy, Model Improvement Incognito chat (ghost icon) Settings, memory
Gemini Google account, Gemini Apps Activity, Keep Activity Keep Activity off Gemini Apps Activity

Menus change. If you can't find a setting, ask the AI itself: "Where do I turn off training on my chats?"

2. Never paste in

Passwords, PINs, account or card numbers, Social Security numbers, or other people's private information. Use placeholders: "Write a letter to [LANDLORD]."

3. Ask well (R.O.L.E.S.)

Role (who it should act as), Objective (what you want), Limits (length, budget, tone), Examples (show what good looks like), Specifics (your situation, plus the actual document or photo).

4. One chat per job

Start a new chat when the topic changes, when it starts forgetting or repeating itself, or when it keeps returning to a wrong idea.

Recap and restart: "Summarize what we've figured out so far as a short recap I can paste into a new chat."

5. Let it think

For multi-step problems (math, cut lists, plans, troubleshooting), turn on the "Thinking" option or ask: "Work through this step by step before answering." Then read the steps. Neat-looking work can still be wrong, and the steps show you where.

6. Starter prompts

7. Stay safe

Free places to start