InnovateED 2026 · October 5

Discovering Claude

There are no slides. Scroll this page as we go. Every prompt is copyable, so you paste as we reach it and leave having done the work rather than watched it.

Two hours, no prior experience assumed. Taught by Nicole Patten.

Open this page on your own screen

elevateonline.com/stanford

claude.ai

Coffee and Claude?

How can I help you today?

0:00

Hello, I’m Nicole

I run Elevate Online and elevateonline.com, and I teach teams how to work with Claude.

Nicole Patten
  • Seven years at Google as a senior UX engineer, on Search and Shopping.
  • Left in October 2025 to do this full time.
  • I run my whole company on Claude: the proposals, the research, the client work, the writing, the code behind this page.
  • Everything I show you today is something I use on a weekday.
  • Before anything else, if you are at a university

    Check whether your institution gives you Claude for Education. Stanford affiliates do, and it is free. It lifts the message limits you will otherwise hit on a free account partway through a session like this one.

  • On a free account, you get a limited number of messages

    Roughly a few dozen in a rolling window, and a slow build of a PDF or a deck spends several. If you run out, Claude tells you when you can start again. Nothing is lost, you just stop for a while.

  • One setting worth changing now

    Settings, then Privacy. On a free account, helping improve the models is on by default. Turn it off if you would rather it was.

Today

  • Set your instructions for Claude.
  • Write a prompt that works.
  • Build a Project for one piece of your work.
  • Skills: repeatable prompts that actually do something. Named today, not built.
0:02In the chat

Rate yourself, and name your task

Four answers in the chat before we start.

One message, like a7 b3 ChatGPT d the accreditation narrative.

Answer in the chat

  1. How comfortable are you with AI in general?1 to 10
  2. How comfortable are you with Claude specifically?1 to 10
  3. If you use AI but not Claude, what do you use? ChatGPT, Copilot, Gemini, whatever it is.Name it
  4. What is the one task that eats more of your week than it should?Name it

In most rooms the first number comes back high and the second one low.

Everything we build today, you build on that last answer. The needs assessment you rewrite every cycle. Summarizing learner evaluations into something a committee will read. Chasing faculty disclosures. The accreditation narrative. An agenda nobody reads.

Start by automating the boring.

0:06

The words you will hear today, and what each one is

Claude is a chat window with some things attached to it. Here is what each one is called, in the order people actually pick them up.

  • A prompt

    What you type into the box. It covers this one task and nothing afterwards. Everything below exists so you stop retyping the same things into every prompt.

  • Instructions for Claude

    One field in Settings that it reads before every message, in every conversation, until you change it. Who you are, how you want things written, what to do when it is unsure. You will write yours today.

  • Memory

    Claude remembering things about you across conversations, managed at Settings and Memory. Different from your instructions: you write those on purpose, and memory accumulates from use. Worth knowing where the switch is before you wonder where something came from.

  • Projects

    A workspace for one body of work, with its own instructions and its own files, so you stop re-explaining the background every time. Free accounts get five. You will make one today.

  • Skills

    A job written down once so it runs the same way every time, by you or by anyone you share it with. Mine turns the week's meals into one shopping list: combined across recipes, grouped by aisle, minus what is already in the cupboard. Every Sunday, same shape, without me retyping any of it. Named today, not built.

  • Connectors

    Claude reading from the systems your documents already live in, rather than from what you paste. Read-only unless you turn writing on, and it only ever sees what you can see. A work account usually needs an administrator to approve it first.

How the three sit together
Instructions for Claudeevery conversation, including both of theseGardeningthe plot planfrost dates for your zoneProjectCookingyour recipe bookwho eats whatProjectA Skillthe week’s shopping list

Your instructions sit above everything and apply first, so both Projects already know who you are. The files are what each Project knows that the other does not: the plot plan and your frost dates on one side, your recipe book and who eats what on the other.

The Skill lives under the Project whose work it does. Mine takes the week’s meals and returns one list, combined across recipes, grouped by aisle, minus what is already in the cupboard. Same job, same shape of answer, every Sunday, which is what makes it a Skill and not a prompt I retype.

Everyone starts in the chat window. What changes over the first few months is how much of your context lives in the other four.

Ask Claude

It will explain its own parts in your terms rather than mine, and draw you a version of the diagram above.

Paste this

I am new to Claude. In plain language, explain the difference between your Instructions, a Project, and a Skill, using the work of someone who runs continuing education at a hospital as the example. Draw it as a simple diagram as well, and tell me which one to set up first and why.
0:10You try it

Write your Instructions for Claude once

One field in Settings. Claude reads it before every message you send.

  • One field, on your account, applying to every conversation until you change it.
  • Who you are, how you want things written, what it should never say.
  • It is a preference, not a contract. Change it the moment it stops fitting.
  • Projects get their own instructions as well, later. This one is the account-wide layer underneath them.

You are not writing these by hand. Open the instructions generator , answer the questions, and paste what it gives you into Claude under Settings. It opens in a new tab so you keep your place here.

Saved? Open a brand new chat and ask it what it now knows. This is the fastest way to see whether what you wrote is actually usable.

In a new chat, once your instructions are saved

What do you now know about me and how I want things written? Tell me which line will change your answers the most, and name anything in there that is too vague for you to act on.

A new chat has nothing in it, so everything it says back came from Settings. The last part is the useful part: whatever it calls vague is a line you wrote for yourself rather than for Claude, and that is the line to go and fix.

On your own8 minutes

Write your instructions and save them in settings

Use the generator, then paste the result into Claude's settings. Start a brand new chat and ask it something you asked earlier today.

You are done when

  • The instructions are pasted into settings and saved.
  • A fresh chat answers in the style you asked for without being told.
  • You can name one line you would change after seeing it work.
0:24You try it

Brief Claude the way you would brief a new colleague

Your instructions handle who you are. A prompt handles what you want this time.

A brilliant new colleague on day one. Sharp, fast, and knows nothing about your work.

Put these three in every prompt

  1. Who is it for, and who should Claude act as?Who
  2. What exactly do you need? Put quotation marks around any wording that has to come back word for word.What
  3. What should come out, and how long should it be?Output

Here is the same request without them and with them.

Before

Make me a bracket for our office tournament.

After

Build a bracket for a 14-team cornhole tournament, double elimination, best of three per match, played Monday through Thursday over two weeks. Make it a PDF with a cool 90s retro style. Use placeholder team names so we can fill in the real ones, and a headline that reads "We came to play!"
  • It named the structure. Without that, Claude picks a format for you and you throw it away.
  • It named the output. A PDF, not something I can print.
  • It gave a visual direction, 90s retro, which removes a whole round of back and forth.
  • It put the headline in quotation marks, so those words come back exactly as written.
  • It said what to do about data that does not exist yet.
On your own6 minutes

Write a full brief for the thing you named

Build the brief below for the thing you need to produce, paste it into Claude, and let it interview you. The questions it asks back are the gaps you left.

You are done when

  • It names the output format, not only the topic.
  • Nothing is left in brackets.
  • You answered what Claude asked rather than letting it guess.

Build your brief

Your prompt

Build [what] for [who], covering [the specific content]. Output it as [file type]. Structure: [sections, length, counts]. Style: [two or three words of direction]. Before you build it, ask me one question at a time until you have what you need.

Anything still in terracotta is a bracket Claude will have to guess at, or ask you about.

0:40You try it

When the answer misses, add the context you left out

The first draft comes back generic and a little stiff. What is missing is almost always something you never said.

  • Do not give up after one flat answer. One try tells you almost nothing.
  • Do not re-ask the same way and hope. If the input does not change, the output will not either.
  • Do not assume it knew context you never gave it.

When the answer misses, you left something out. The tool is almost never the problem.

Ask Claude

Before you rewrite a prompt yourself, make it tell you what is missing.

Paste this

Here is a prompt I wrote. Do not answer it. Instead, list what you would still have to guess at, and ask me the three questions that would most change your answer.

Before

This is not quite right, can you make it better?

After

Can you add some graphics or diagrams? Also the dark blue should be an accent color, since this needs to print we want a light background. I can also see some of the tips bleeding over the edge of the boxes.
  • It named something missing. Better does not point at an absence.
  • It gave the reason behind a change, because it has to print. The reason is worth more than the instruction, since it lets Claude apply the same logic to things you did not mention.
  • It pointed at a defect you can see and Claude cannot.

A retry template, if you want one.

A retry template

Three things to fix:
1. It is missing [the thing that is not there].
2. [This part] is wrong because [why].
3. Change [this] to [that], because [the reason].

Keep the structure, change only what I named.
On your own5 minutes

Name three specific defects, then retry

Take the output you just made and write a retry that names one thing missing, one thing wrong, and one reason behind a change.

You are done when

  • Your retry names something absent, not only things you dislike.
  • At least one instruction carries its reason.
  • You ran it, and the second answer is closer.
0:53You try it

Claude builds files you can open and send

The output does not have to stay in the chat window. Claude builds the file and hands it to you.

  • Documents, slide decks, spreadsheets, printable one-pagers.
  • Charts and diagrams, from numbers you give it or from a process you describe.
  • Analysis that reads a file you upload and works from what is in it.
  • Not images. That is the one thing to go elsewhere for, then bring the result back here.

Pick a topic you already know cold, so you are judging the file and not the facts. How to make espresso. The rules of pickleball. Your dog. Why sourdough works. The guitar you never learned to play. Anything you could talk about for five minutes without looking something up.

Run the first one, open what comes back, then run the second in the same chat. The second only works because the first is still in the conversation.

First

Create a five slide presentation about [your topic].

Then, in the same chat

Now create a printable one page version of the same thing, written for someone brand new to it.

A build like that one can take a few minutes. While it runs, do not sit and watch it. Open a second chat and try this, which is about you rather than about a topic.

Your week, as a diagram

Create a diagram of the typical weekly workflow for a [your role]. Show where the time actually goes and where the handoffs to other people are. Then mark the three steps most worth handing to AI, and say why for each.

Read what comes back against your own week. What it gets right is what is typical of the role. What it misses is usually specific to you.

On your own5 minutes

Make the same content as two different files

Ask for your topic twice, the two ways above, and open both files.

You are done when

  • You have two files open, not two blocks of text in the chat.
  • You can name one thing the deck does that the printable does not.
  • You changed who it is for on the second ask and can see where that landed.
1:02Halfway

Check what stuck before we change subject

Four questions on the first hour. Nothing is recorded and nobody sees your answers. Click one and the reason shows up either way.

Your first draft comes back generic and a little stiff. What is most likely going on?

You wrote your Instructions for Claude this morning. Which of these belongs in them?

You want a one-page printable you can hand out, not text in the chat. When do you say so?

Claude gives you a confident answer with a date and a citation in it. What now?

1:05

Long conversations drift, so start a new one

Claude holds the whole conversation at once: everything you typed, everything it said back, everything you pasted. Once that fills up, something has to give.

  • It is not forgetting out of carelessness. It has run out of room.
  • When a chat starts contradicting itself or dragging in a topic from an hour ago, that is the signal.
  • A fresh chat is not starting over. It is clearing the noise so it can focus on the thing in front of it.
Everything it is holding at once

Every message you sent, every answer it gave, and every file you pasted sits in one space, and Claude re-reads all of it before each reply.

As that space fills, the oldest parts get compressed rather than deleted. Nothing errors out. The detail at the top just stops being sharp, which is why a long chat starts contradicting itself instead of announcing a problem.

Stay in the chat you are in

  • You are still working on the same piece. Drafting, then fixing, then fixing again.
  • It needs what you already told it. The file you uploaded, the decisions you made twenty messages ago.
  • You are asking follow-ups. Shorter, cheaper and better than re-explaining.

Start a new one

  • You have moved to a different subject. A new chat costs nothing.
  • It is contradicting itself, or dragging back something you settled an hour ago.
  • It has started ignoring an instruction it was following earlier.
  • You are about to paste something confidential into a chat that has unrelated history in it.

Start the new chat. Carry three sentences, not the whole history.

  • Name your threads

    Claude titles them for you and the titles drift. Rename the ones you will come back to, because each thread keeps its own context and you will want to find the right one.

  • Incognito, if you want a chat that leaves no trace

    The ghost icon at the top right. That chat does not go into your history and nothing from it is remembered, the same way an incognito browser window works. It is not the same thing as turning off model training, which is a separate switch in Settings under Privacy.

  • Memory is a setting, and it is yours to edit

    Settings, then Memory. It builds up from your conversations, so the more you work with Claude the more it knows. You can read what is in there and tell Claude to forget any of it.

When you are not sure, ask it.

Ask it directly

How full is your context window right now, and would this conversation benefit from starting fresh?
1:12You try it

Claude states wrong answers confidently, so check sources

Claude states a wrong answer in exactly the same tone as a right one. Nothing in how it sounds tells you which one you are reading.

  • Ask for the source. Then check the source yourself, rather than checking that a source was named.
  • Anything consequential goes past a human who knows the subject.

Confident phrasing is not evidence. Every number, date, name and citation needs a source you checked yourself.

The verification prompt

List every factual claim in this draft that you are not fully certain about, including any numbers, dates, names, or citations, and tell me exactly where I should verify each one.

Better still, ask for the sourcing up front rather than auditing it afterwards. Use something you genuinely need answered, not a test question, because the point is to see whether you can act on what comes back.

Research something, with the receipts

Research [a question you actually need answered] and give me the three findings most relevant to my work. Under each one, tell me what it is based on and how confident you are. Where you cannot point to a specific source, say so instead of naming one. I would rather have a gap than a guess.

That last line is the one that matters. Asked for sources without it, a model will sometimes produce a citation that looks right and is not, which is the same failure in a smarter costume.

On your own5 minutes

Run the source check on your own output

Take anything Claude produced for you in the last hour and run the prompt above on it. Read what it flags.

You are done when

  • You have a list of claims it is not sure about.
  • You picked one and went and checked it against the source.
  • You can say whether the list was useful or over-cautious.
1:22

Sort what you paste: goes in, redact, stays out

One minute on this, because it comes up and then it is settled.

  • Goes in. Your own drafts, your own notes, your own slides, anything already public. This is most of it.
  • Goes in redacted. The same document with the two identifying lines taken out.
  • Stays out. Other people's confidential material: a learner's evaluation with their name on it, anything carrying patient detail, anything under an agreement you signed.

If you would not want it quoted with your institution’s name attached, do not paste it.

Ask Claude

It will give you the general rule. Your own policy still outranks it.

Paste this

I work in continuing education at a hospital. What kinds of information should I avoid pasting into you, and what is usually safe? Where my employer's own policy would decide it rather than you, say so instead of guessing.

The common mistake is not over-sharing, it is holding back work that was always fine to use. Your own drafts, your agendas, published guidance and anything already public all go in as they are. Where your institution’s policy or your privacy office decides it rather than you, the useful answer is knowing who to ask.

1:44You try it

Put work you repeat into a Project

You would not keep your garden notes and your recipes in the same notebook. Not because it would break, but because you would have to say which one you meant every single time.

  • A chat is a conversation you have once. A Project is a desk you come back to, with the files already on it.
  • It holds its own instructions and its own files, so you stop re-explaining the background every time.
  • Free accounts can have up to five. That is more than enough to find out whether you like working this way.

True of you in every conversation? Instructions. True of one body of work? Project.

  • Account instructions: who you are, everywhere.
  • A Project: what one body of work needs.

Projects people in this room actually make

  • Accreditation. The criteria, your policies, last cycle’s self-study and the reviewer feedback you got back.
  • One activity or conference. The agenda, the objectives, the faculty list, last year’s evaluations.
  • Needs assessment. Gap data, survey results and the literature you keep citing, so the narrative comes out the same shape every cycle.
  • Faculty and disclosures. The policy, the forms, and the chasing emails you send forty times a year.
  • Evaluations and outcomes. The exports, summarized the same way each round so two rounds are comparable.

Its instructions are two or three lines about the work, not about you. What is always true of it, who it is for, and the shape you almost always want back. The files you upload sit underneath, and Claude reads them before it asks you for background.

On your own8 minutes

Create a Project for the thing you named

Make a Project named after the thing you put in the chat at the start, give it two or three lines about that work, and upload one document it would always need.

You are done when

  • The Project exists and is named after real work, not Test.
  • Its instructions say something true of the work, not of you.
  • You asked it something and the answer used the file without you attaching it.
1:50Recap

You built three of the four

The same four from the start, in the same order, and what you now have of each.

  • Your instructions for Claude

    Who you are, how you want things written, what to do when it does not know. Written once, in Settings, applying to every conversation from now on. Yours are saved.

  • Prompting

    Name the output, the structure, the style, and quote any wording that has to come back word for word. When the answer misses, the missing context is where you fix it, not the tool. You built one and ran it.

  • Projects

    A workspace with its own instructions and its own files, so you stop re-explaining the background. You have one, named after real work, with a document in it.

  • Skills

    Named today, not built. Go looking for this one once the first three are habit.

Three settings and a habit. That is the whole operating layer, and you already have it.

1:55

Where to go from here

You are two rungs up a four-rung ladder. Here is the rest of it, named so you know what to search for. The appendix at the bottom has the detail and some things to try tonight.

  • You have the first two: instructions that travel with you, and a Project holding one body of work.
  • Next is a Skill. The job written down once so it runs the same way every time, and so you can hand it to someone else.
  • Then connectors, so it reads from where your material already lives. Over two hundred of them, and Instacart is one, which tells you how far past work this goes.
  • Then scheduled work: the Skill running on its own on a Monday morning, with the output waiting for you.
AppendixNot taught today

Things worth trying tonight

Nothing here is on the clock. Every one works on a free account unless it says otherwise.

  • Photograph your handwriting and get a document back

    A whiteboard, a page of notes, a receipt. Claude reads the image and turns it into something you can edit. This is the one that gets the loudest reaction in a room, and it costs nothing to try.

  • Hand it a mess and ask for a table

    A pasted list, an email thread, a page of numbers. Ask for it as a table, then ask for a chart of the same thing. The chart comes back as an interactive artifact you can open beside the chat.

  • Make something that runs, not something that reads

    A flashcard set for what you are studying, a quiz on a reading, a small calculator for a decision you keep making. Ask for it as an interactive artifact.

  • Bring over what ChatGPT already knows about you

    Ask ChatGPT to write a summary of what it has learned about how you work and what you want from it, then paste that into your Instructions for Claude. There is a longer walkthrough at switchtoclaude.com.

  • Make it argue with you

    Give it your own position and ask for the strongest case against it, then ask what evidence would change your mind. This is the use that holds up best against a skeptical room.

  • Upload a document you already know well

    A guideline, a report, a paper you have read twice. Ask for the three claims its author would defend hardest and the one a reviewer would go after first. You are checking its reasoning against yours, which is the only way to get calibrated on where it is weak.

Watch, do not type

These need a connected account, so they are a demo rather than an exercise. They are here so you can find them again, and so you can see what the version of this with your own systems attached actually looks like.

Across a week of mail

Look at my email from the last seven days. What did I say I would do and have not done yet? Give it to me as a list, most overdue first, with who is waiting on each one.

Before a meeting

I have a meeting with [name] at [time]. Pull what we last talked about from my mail and calendar, and give me the three things I should walk in knowing. Say if you cannot find anything rather than guessing.

Across a folder of documents

Read the documents in [folder] and tell me what question they all seem to be circling but none of them answers directly.

Three more worth knowing exist, because they are what people mean when they say Claude did the whole thing. Connectors are on every plan, including the free one you set up this morning.

  • Connectors, so it works on live things instead of pasted text

    Claude reads from where your material already lives. The directory has over two hundred, and not all of them are work: Instacart will build a grocery cart from a recipe, with live inventory from stores near you. Read-only by default, writing is an opt-in, and a connector only ever sees what you can see. A work account may need an administrator to approve it first.

  • Scheduled work that runs while you are not there

    A job that runs on a schedule and leaves the output waiting for you. This lives in Claude's desktop app rather than the browser.

  • What a paid account changes

    More messages, the better models, and Cowork. On Team and above, Anthropic does not train on your data at all: it is contractual rather than a setting someone could flip. If you handle anything sensitive, that difference is the whole argument.

  • Claude Code, if you write any

    The same engine, pointed at a repository instead of a chat window. Worth a look if you have ever maintained a script, and not only for software engineers.

Use Claude on one weekly task, five times

  • Five runs on work you already understand is how you learn the shape of what Claude is good at.
  • On unfamiliar work you cannot tell a good answer from a confident one.
  • This page stays up. Bookmark it and take the prompts with you.

If your team wants to go further together, the work we do is at elevateonline.com. No obligation, and no follow-up unless you ask for one.