Skip to content
Leadership Consultant Studio · Free sample Members: sign in to the full Studio

Session 1: What Claude Actually Does

By the end of this session you can explain why a fluent answer is not proof, open Claude and hold a safe first conversation, and turn a vague request into a framed assignment whose reply you can check.

A reading copy of Session 1 from the free sample. In the members’ Studio, the same session has checklists, notes, self-ratings and Mentor Mode, and keeps your place.

How this session works: three phases

  1. Phase 1: Prepare for Class

    Pre-work on your own, before Learn Together.

    Effort: about 10 to 15 minutes.

  2. Phase 2: Learn Together

    Five stages, with your mentor or on your own.

    Effort: about 48 minutes with a mentor, about 65 to 75 minutes on your own.

  3. Phase 3: Apply Independently

    Practice on your own, after Learn Together.

    Effort: about 30 to 40 minutes for the required exercises, plus about 15 minutes for the optional exercise 3.

Session 1: Prepare for Class

Phase 1 of 3. Effort: about 10 to 15 minutes. These are estimates, not measured times. You do not need to finish this before the live session.

What you will be able to do

Explain in plain language why Claude can give a persuasive answer that is still wrong, and open Claude for a first safe exchange.

Concept primer

What a language model does

A language model is a computer program trained on a very large amount of text. From that training it learned patterns: which words, ideas and ways of arguing tend to follow which.

When you send it a message, it writes a reply a small piece at a time. At each step it chooses a likely next word, or part of a word, given your message and everything earlier in the conversation. It does this very well, which is why the replies read so fluently.

It is not looking up a stored, checked answer. Unless you give it documents, or it has a tool that searches the web, it works only from patterns learned in training and from what you have told it.

Its training stops at a cutoff date. On its own it may not know about anything that happened after that.

How a reply gets writtenThe model reads your message and the conversation, predicts a likely next piece of text, adds it and repeats until the reply is complete. The result is fluent, but it is predicted, not looked up.
  1. Your message and the conversation so farPlus anything you attached.
  2. Predict a likely next pieceFrom patterns learned in training, shaped by what is in front of it.
  3. Add it, then repeatWord by word, until the reply is complete.
  4. Your jobA fluent replyWritten, not retrieved. Checking it is your job.
An analogy

Picture a very well-read colleague who has read an enormous library but carries no notes. Ask a question and they answer fluently from their memory of everything they have read, shaped by exactly how you asked.

Where the analogy breaks down: A colleague usually knows when they are unsure and says so. A language model does not reliably know what it does not know, and it cannot go back to a particular book to check. It also has no understanding of your situation beyond the words in front of it.

Common misconceptions
  • It looks the answer up. It generates text from learned patterns. Nothing in the reply has been checked against a source unless you supplied one.
  • If it sounds sure, it is right. A confident tone is part of the writing, not evidence. A fluent answer can be wrong.
  • It knows my client. It knows only what is in the conversation and what it learned in training. Anything specific about your client either came from you or was guessed.
  • It remembers everything I have told it before. What it can draw on depends on the conversation and the product's settings. Session 2 covers this.

Try it: Context quality: vague, framed, and grounded (members’ Studio) (illustrative)

Fluent is not the same as true

Two questions are easy to blur: is this answer well written, and is it correct? A language model is very good at the first. The second one is your job.

Because the model writes the most plausible continuation, it can produce statements that sound exactly right but have no support: a statistic nobody measured, a source that does not exist, or a detail about a client you never gave it. People often call this a hallucination.

The risk is highest where the model has to fill a gap: when the request is vague, when it is asked about specifics you did not supply, and when it is asked for names, dates, figures, quotes or citations.

The habit that protects you is simple. For every statement that matters, ask where it came from: from what you supplied, from reasonable inference, or from nowhere you can point to.

Two separate questionsA two-by-two grid. How a reply reads (fluent or clumsy) and whether it holds (supported or unsupported) are independent. The dangerous corner is fluent and unsupported, because it sounds right.

Across: how it reads. Down: whether it holds.

Fluent Clumsy
Supported Fine, once you have checked it. Fix the wording; the content is sound.
Unsupported The dangerous one: it sounds right, so it slips through. Easy to catch, because it already looks wrong.
An analogy

A polished slide deck from a confident new hire. The structure is good and it looks finished, but you would still check the numbers before it went to a client.

Where the analogy breaks down: A new hire who invents a figure usually knows they did. A language model does not intend to mislead and gives no signal that a detail was filled in, so tone and hesitation will not help you spot it.

Common misconceptions
  • It only gets hard questions wrong. It can be wrong about simple specifics too, especially names, dates, figures and sources.
  • Asking it whether it is sure checks the answer. Asking it to review its own answer can surface problems, but it is still the model checking itself. Verifying means comparing with a source you trust.
  • A well-organized answer is a reliable answer. Structure and accuracy are separate. You can keep a good structure and still need to check every claim inside it.

Try it: Context quality: vague, framed, and grounded (members’ Studio) (illustrative)

Optional: Concept lab: Context quality: vague, framed, and grounded (members’ Studio). Illustrative example. These replies are static text written for teaching. They are not live responses from an AI model, and a real reply will differ.

Meet the Claude feature

Open Claude, find where you type and where past conversations are listed, and send one safe test message. If the desktop app is not available, use Claude in a web browser.

Toolkit: Claude Desktop: install, sign in and find your way around (members’ Studio)

Readiness check

You will need:

If something does not work: If Claude Desktop will not install or open, use Claude in a web browser at claude.ai. Everything in Session 1 works there. Make a note of what went wrong, to raise with your mentor or to come back to after the session, and do not spend more than a few minutes on it. Source: Get started with Claude (opens in a new window) (Claude Help Center, checked 2026-11-20)

Bring to Learn Together

One question about how Claude works, or one setup problem you ran into, to raise with your mentor or to come back to as you work through the session. If everything worked, keep the test message you sent and what you noticed about the reply.

Session 1: Learn Together

Phase 2 of 3. Five stages. Effort: about 48 minutes with a mentor, about 65 to 75 minutes on your own. These are estimates, not measured times. Timing is flexible.

Five stages

  1. Connect and Review (About 5 minutes)
  2. Understand the Technology (About 10 minutes)
  3. Demonstrate in Claude (About 13 minutes)
  4. Guided Practice (About 15 minutes)
  5. Reflect and Assign (About 5 minutes)

Stage 1 of 5: Connect and Review

Session 1, Learn Together. About 5 minutes.

Welcome to the first session. In it you will see what Claude actually does, see it work on a request, and practice turning a vague request into one you can check.

Before we start

  • Is Claude open and signed in, in the desktop app or a web browser?
  • What did you try in the pre-work, and did anything get in the way?
  • Your reflection: what makes a fluent, confident answer seem trustworthy?

On your ownIf you are working alone, answer these for yourself in a few words before you go on. If something from the pre-work is still not working, the Toolkit tutorial below covers the browser route.

Opening Claude, on the desktop or in a browser: Toolkit: Claude Desktop: install, sign in and find your way around (members’ Studio)

Ground rules for every session

  • Use general or fictional content only. Our fictional client is Brindlecrest Health Partners.
  • Nothing Claude produces goes to anyone until you have checked it.
  • Questions are welcome at any point, including "I don't know what I'm looking at." If you are working alone, note each question as it comes up and look for it in the Toolkit.

Today's fictional client wrote this:

Engagement inquiry from Brindlecrest Health Partners

FICTIONAL TRAINING MATERIAL

Email from the chief executive to the consultant. Brindlecrest Health Partners, the Larkmoor region and everyone named are invented. Figures marked with a dagger are made up for training.

From: Corrin Ashdale, Chief Executive, Brindlecrest Health Partners (fictional)

To: the consultant

Subject: Help with our executive team

I lead Brindlecrest Health Partners, a community healthcare organization, and have been its chief executive for 2 (synthetic figure) years. [INQ-01] Last year we restructured from 2 (synthetic figure) regional divisions into a single operating model for the whole organization. [INQ-02]

Since the restructure, my executive team of 7 (synthetic figure) people has not been working well together. [INQ-03] Decisions we agree in our weekly executive meeting are often reopened afterwards in one-to-one conversations. [INQ-04] Our meetings have become tense, and some executives say very little in them. [INQ-05] I have heard second-hand that some executives feel they were not consulted about the restructure. [INQ-06]

I believe the core problem is communication. [INQ-07] I would like you to run a 1-day (synthetic figure) communication skills workshop for the executive team. [INQ-08] We would like the workshop to happen within the next 3 (synthetic figure) months. [INQ-09]

Could we arrange a 45-minute (synthetic figure) call to discuss how you might help? [INQ-10] Please treat this message as confidential. [INQ-11]

Corrin Ashdale

The copy is plain text: the banner, the title, the document and its claim numbers.

Open as its own page (members’ Studio)

Check yourself

Before you move on, answer these for yourself. With a mentor, talk them through.

  • What did you try in the pre-work, and what happened?
  • What do you think happens when you send Claude a message?
  • Looking at your reflection, what made a confident answer feel trustworthy?

Checkpoint

Claude is open in front of you, and you know today's goal: understand why a fluent answer is not proof, and practice framing a request so its reply can be checked.

Stage 2 of 5: Understand the Technology

Session 1, Learn Together. About 10 minutes.

Two ideas carry this whole course. Read them, then put them in your own words, out loud to your mentor or in your notes for this session.

What a language model does

  • It writes a reply one likely piece at a time, from patterns learned in training and from what you give it.
  • It does not look up a checked answer, and its training stops at a cutoff date.
  • Fluent writing is how it works, not a sign that it is right.
Remind me: the full explanation
What a language model does

A language model is a computer program trained on a very large amount of text. From that training it learned patterns: which words, ideas and ways of arguing tend to follow which.

When you send it a message, it writes a reply a small piece at a time. At each step it chooses a likely next word, or part of a word, given your message and everything earlier in the conversation. It does this very well, which is why the replies read so fluently.

It is not looking up a stored, checked answer. Unless you give it documents, or it has a tool that searches the web, it works only from patterns learned in training and from what you have told it.

Its training stops at a cutoff date. On its own it may not know about anything that happened after that.

How a reply gets writtenThe model reads your message and the conversation, predicts a likely next piece of text, adds it and repeats until the reply is complete. The result is fluent, but it is predicted, not looked up.
  1. Your message and the conversation so farPlus anything you attached.
  2. Predict a likely next pieceFrom patterns learned in training, shaped by what is in front of it.
  3. Add it, then repeatWord by word, until the reply is complete.
  4. Your jobA fluent replyWritten, not retrieved. Checking it is your job.
An analogy

Picture a very well-read colleague who has read an enormous library but carries no notes. Ask a question and they answer fluently from their memory of everything they have read, shaped by exactly how you asked.

Where the analogy breaks down: A colleague usually knows when they are unsure and says so. A language model does not reliably know what it does not know, and it cannot go back to a particular book to check. It also has no understanding of your situation beyond the words in front of it.

Common misconceptions
  • It looks the answer up. It generates text from learned patterns. Nothing in the reply has been checked against a source unless you supplied one.
  • If it sounds sure, it is right. A confident tone is part of the writing, not evidence. A fluent answer can be wrong.
  • It knows my client. It knows only what is in the conversation and what it learned in training. Anything specific about your client either came from you or was guessed.
  • It remembers everything I have told it before. What it can draw on depends on the conversation and the product's settings. Session 2 covers this.

Try it: Context quality: vague, framed, and grounded (members’ Studio) (illustrative)

Fluent is not the same as true

  • Well written and correct are two separate questions; the second is your job.
  • The risk is highest where it has to fill a gap: names, figures, quotes, sources, details you never gave it.
  • For every statement that matters, ask where it came from.
Remind me: the full explanation
Fluent is not the same as true

Two questions are easy to blur: is this answer well written, and is it correct? A language model is very good at the first. The second one is your job.

Because the model writes the most plausible continuation, it can produce statements that sound exactly right but have no support: a statistic nobody measured, a source that does not exist, or a detail about a client you never gave it. People often call this a hallucination.

The risk is highest where the model has to fill a gap: when the request is vague, when it is asked about specifics you did not supply, and when it is asked for names, dates, figures, quotes or citations.

The habit that protects you is simple. For every statement that matters, ask where it came from: from what you supplied, from reasonable inference, or from nowhere you can point to.

Two separate questionsA two-by-two grid. How a reply reads (fluent or clumsy) and whether it holds (supported or unsupported) are independent. The dangerous corner is fluent and unsupported, because it sounds right.

Across: how it reads. Down: whether it holds.

Fluent Clumsy
Supported Fine, once you have checked it. Fix the wording; the content is sound.
Unsupported The dangerous one: it sounds right, so it slips through. Easy to catch, because it already looks wrong.
An analogy

A polished slide deck from a confident new hire. The structure is good and it looks finished, but you would still check the numbers before it went to a client.

Where the analogy breaks down: A new hire who invents a figure usually knows they did. A language model does not intend to mislead and gives no signal that a detail was filled in, so tone and hesitation will not help you spot it.

Common misconceptions
  • It only gets hard questions wrong. It can be wrong about simple specifics too, especially names, dates, figures and sources.
  • Asking it whether it is sure checks the answer. Asking it to review its own answer can surface problems, but it is still the model checking itself. Verifying means comparing with a source you trust.
  • A well-organized answer is a reliable answer. Structure and accuracy are separate. You can keep a good structure and still need to check every claim inside it.

Try it: Context quality: vague, framed, and grounded (members’ Studio) (illustrative)

See how the request changes the reply (illustrative examples, not live responses): Concept lab: Context quality: vague, framed, and grounded (members’ Studio)

On your ownIf you are working alone, open the lab and compare the vague and framed examples. In the framed example, find one thing the reply assumed without being told.

Three things to remember

  1. Claude writes a likely reply from learned patterns. It is not retrieving a checked answer.
  2. Well written and correct are separate questions, and you answer the second one.
  3. The fewer gaps you leave, the less it has to guess.

Check yourself

Before you move on, answer these for yourself. With a mentor, talk them through.

  • If Claude is not looking things up, where does its answer come from?
  • Why might a confident tone be misleading?
  • In your own work, where would a plausible but invented detail do the most damage?

Checkpoint

In one or two sentences, explain to your mentor why a confident answer from Claude can still be wrong.

On your ownWrite those one or two sentences in your own words without looking back, then reread the two concept cards and see what you would change.

Stage 3 of 5: Demonstrate in Claude

Session 1, Learn Together. About 13 minutes.

In this stage you see Claude work on a deliberately vague request: your mentor runs it, or you run it yourself. Watch three things: what Claude does with what it was told, what it adds on its own, and how sure it sounds.

Minimum to finish today: demonstration 1. Demonstration 2 is if you have time.

On your ownIf you are working alone, run the demonstration yourself with the request on this page:

  1. Open Claude and start a new chat, as the chat tutorial below describes.
  2. Copy the demonstration request and send it exactly as it is. Do not add any detail: the gaps are the point.
  3. Read the whole reply once. Then go through it again and note each detail Claude could not have known from the one line you sent.
  4. Find the three places listed under "Where things are". Leave the model as it is.
  5. Find this conversation again in your list of past conversations.

Where things are

Step by step, with the Windows and Mac differences: Toolkit: Claude Desktop: install, sign in and find your way around (members’ Studio)

Sending messages, choosing a model and attaching files: Toolkit: Claude chat and model selection

While you watch or run it, note

  • One thing in the reply that came from the request.
  • One thing Claude supplied that the request never mentioned.
  • Any place it sounded certain about something it could not have known.

Demonstration 2, if you have time: a request that invites invented sources (about 3 minutes)

The first request was vague, so Claude may have answered in general terms or asked a question. This one asks for exactly the kind of detail that is easy to invent.

  1. Start a new chat and send the second demonstration request below.
  2. Choose one study from the reply and look it up yourself, in a web browser or a library search.
  3. Record whether it exists, whether the authors, year and journal match, and whether the finding matches what the study says.

Claude may decline, hedge, or search the web if search is on. Each of these is a useful result. If it returns linked sources, check that each source says what the line claims.

Prompts for this stage

The demonstration request (deliberately vague)

Help me prepare for a coaching conversation with an executive who is struggling with their team.

Demonstration 2: a request that invites invented sources

List three published studies, with authors, year and journal, showing that trust within an executive team improves its performance. Give a one-line finding for each.

Check yourself

Before you move on, answer these for yourself. With a mentor, talk them through.

  • Which details in the reply came from the request?
  • What did Claude add on its own? Is any of it about a real person or situation?
  • How sure did it sound, and does that match how much it was told?
  • Did the reply give you anything you could not find, or a real source that says something different?

Checkpoint

You can point to where you type, where the reply appears, which model is answering, and where this conversation is listed.

Stage 4 of 5: Guided Practice

Session 1, Learn Together. About 15 minutes.

Now you direct the work. The method is the operator framework: Frame, Ground, Verify, Govern. Today you practice Frame and Verify, with a first taste of Ground and Govern.

Remind me: Frame

Frame: the 7 parts of an assignment

Decide what you want before you ask, and write it down as an assignment, not a wish.

Objective
What the work is for, and what decision or deliverable it serves.
Role
Who the assistant should act as, and whose point of view matters.
Context
The background it needs, which you supply. It knows nothing about your client unless you tell it.
Assumptions
What you are taking as given. Saying them out loud lets you and the assistant question them.
Boundaries
What to avoid, what material to stay within, what is confidential, and what to do when something is not known.
Output format
The shape you need back: length, structure, audience and tone.
Review criteria
How you will judge the result, decided before you see it.

The whole framework: Frame, Ground, Verify, Govern (members’ Studio)

Your task

  1. Start a new chat.
  2. Copy the Frame template below and paste it into the chat. Do not send yet.
  3. Fill in the parts in square brackets. Leave the last one, the inquiry placeholder, for the next step.
  4. Ground the request: open the Brindlecrest engagement inquiry (the case document below), copy its text, and paste it in place of "[Paste the Brindlecrest engagement inquiry here]".
  5. Send it, and read the whole reply.
  6. Verify: go through the reply line by line. For each specific statement, decide whether it came from the inquiry, is a reasonable inference, or is unsupported.
  7. Challenge: send the follow-up prompt below, then compare Claude's list with your own.
  8. Govern: decide which questions you would actually take into the conversation, and what you would check first.

On your ownIf you are working alone and a Frame part is hard to write, write one plain line for it and move on. Note which part it was, and look at what the reply did with it.

The four habits you will use in every session. Today you practice Frame and Verify.

Frame, Ground, Verify, GovernThe four habits in order: write a clear brief, supply the material to work from, check every statement that matters, then decide, approve and record. What you record feeds the next piece of work.
  1. FrameWrite the brief: objective, role, context, assumptions, boundaries, output format, review criteria.
  2. GroundSupply the material to work from, and say to stay within it.
  3. VerifyCheck each statement that matters against the source. Challenge the result.
  4. You decideGovernDecide, approve and record. You stay accountable.

What you record under Govern improves the next Frame.

Source material:

Engagement inquiry from Brindlecrest Health Partners

FICTIONAL TRAINING MATERIAL

Email from the chief executive to the consultant. Brindlecrest Health Partners, the Larkmoor region and everyone named are invented. Figures marked with a dagger are made up for training.

From: Corrin Ashdale, Chief Executive, Brindlecrest Health Partners (fictional)

To: the consultant

Subject: Help with our executive team

I lead Brindlecrest Health Partners, a community healthcare organization, and have been its chief executive for 2 (synthetic figure) years. [INQ-01] Last year we restructured from 2 (synthetic figure) regional divisions into a single operating model for the whole organization. [INQ-02]

Since the restructure, my executive team of 7 (synthetic figure) people has not been working well together. [INQ-03] Decisions we agree in our weekly executive meeting are often reopened afterwards in one-to-one conversations. [INQ-04] Our meetings have become tense, and some executives say very little in them. [INQ-05] I have heard second-hand that some executives feel they were not consulted about the restructure. [INQ-06]

I believe the core problem is communication. [INQ-07] I would like you to run a 1-day (synthetic figure) communication skills workshop for the executive team. [INQ-08] We would like the workshop to happen within the next 3 (synthetic figure) months. [INQ-09]

Could we arrange a 45-minute (synthetic figure) call to discuss how you might help? [INQ-10] Please treat this message as confidential. [INQ-11]

Corrin Ashdale

The copy is plain text: the banner, the title, the document and its claim numbers.

Open as its own page (members’ Studio)

If Claude's own check misses something you found, that is the lesson of today. Asking Claude to review its answer can help you find problems, but it is not verification. Verification means checking against a source you trust.

Prompts for this stage

Frame template (fill in the parts in square brackets)

Objective: Help me prepare for my first conversation with the chief executive of Brindlecrest Health Partners, who has asked for help with their executive team.
Role: Act as a thinking partner for an experienced leadership consultant. [Change this if you want a different role.]
Context: The chief executive's inquiry email is pasted below. It is fictional training material.
Assumptions: [What are you taking as given? For example, that this is an exploratory conversation, not a sales meeting.]
Boundaries: Use only the inquiry below. Do not invent facts about the organization or its people. If something is not stated, say so.
Output format: [For example: open questions grouped by theme, short enough to ask aloud.]
Review criteria: [How will you judge the result? For example: every question can be traced to the inquiry, or is clearly marked as exploratory.]

Inquiry:
[Paste the Brindlecrest engagement inquiry here]

Optional, for everyday requests later: a three-line Frame

Goal: [what this is for, and who will read it]
Material: [what to work from] If something is not stated, say so.
Good looks like: [length and format, and how you will check it]

Use the full seven-part Frame when the work matters or will reach a client.

Challenge follow-up

Go back through your answer. List each statement that is not supported by the inquiry I gave you. For each one, say whether it is a reasonable inference or an assumption with no basis in the inquiry.

Check yourself

Before you move on, answer these for yourself. With a mentor, talk them through.

  • Which Frame part was hardest to write, and why?
  • Can you point to one statement in the reply and say where it came from?
  • What did the challenge prompt find that you missed, and what did you find that it missed?
  • Which of these questions would you actually ask, and which would you drop?

Checkpoint

You sent a framed assignment, checked every specific statement in the reply, and found at least one that the inquiry does not support, or confirmed, after checking each one, that there were none.

Stage 5 of 5: Reflect and Assign

Session 1, Learn Together. About 5 minutes.

Before you finish, answer these three questions, out loud with your mentor or in writing in your reflection.

  1. What did Claude actually know today, and what did you give it?
  2. What in today's replies could have been wrong, and how would you have found out?
  3. What would you need to check or approve before anything like this reached a client?

Rate yourself

On the session overview, rate yourself on Concept, Tool, Operator and Safety. These ratings are for your own reflection. Emerging is a normal place to start.

Before Session 2

Apply Independently has three short exercises that repeat today's method on your own non-confidential tasks.

  • Keep one before-and-after pair: a vague request and your framed version, with no client names or confidential details.
  • Keep one statement from a reply that you could not verify.

Write both in your Apply reflection when you finish the exercises. It appears at the start of Session 2, whether you go on with a mentor or on your own.

Check yourself

Before you move on, answer these for yourself. With a mentor, talk them through.

  • For each of the three questions above, can you name one specific example from today?
  • How would you explain your choice of questions for the chief executive without pointing to Claude as the authority?
  • Which part of today's method do you most want to practice in the homework?

Checkpoint

You know what the homework is and what to carry into Session 2.

Session 1: Apply Independently

Phase 3 of 3. Effort: about 30 to 40 minutes for the required exercises, plus about 15 minutes for the optional exercise 3. These are estimates, not measured times.

Two exercises and an optional third, each a little more independent than the last. Use your own everyday work, kept general or fictional, so you practice on the kind of request you really make.

Use general or fictional content only. Do not enter client names, coaching notes or personal information.

Exercise 1: Guided (about 15 to 20 minutes)

Pick a non-confidential task from your own work, such as a welcome note for a workshop, an agenda for a team offsite, or questions for a first client call. Keep the details general or fictional.

  1. Ask Claude for it in one line, the way you normally would. Save the request and the reply in your notes or a plain document.
  2. Start a new chat. Ask for the same thing using the Frame template. Save the request and the reply in the same place. Your best pair goes in your Apply reflection below, where Session 2 reads it.
  3. Compare the two replies. In each, mark one statement you could not trace to anything you supplied.

Frame template for your own task

Objective: [What is this for? What decision or deliverable does it serve?]
Role: [Who should Claude act as?]
Context: [The background Claude needs. General or fictional only.]
Assumptions: [What are you taking as given?]
Boundaries: [What to avoid, what to stay within. Add: "If something is not stated, say so."]
Output format: [Length, structure, audience, tone.]
Review criteria: [How will you judge the result?]

Task: [The request itself.]

Exercise 2: Supported (about 15 to 20 minutes)

Choose a second task. This time write the framed request yourself, without the template, using the seven Frame parts from memory.

  1. Send it and read the whole reply.
  2. Check it yourself first: for each specific statement, decide whether it came from what you supplied, is a reasonable inference, or is unsupported.
  3. Then send the challenge follow-up and compare Claude's list with yours.

Challenge follow-up

Go back through your answer. List each statement that is not supported by what I gave you. For each one, say whether it is a reasonable inference or an assumption with no basis in what I provided.

Exercise 3: Independent (optional, about 15 minutes)

Choose a third task yourself. Decide which Frame parts matter most for it, write the request, and run it.

Then act as the operator: write down what you would use from the reply, what you would change, and what you would have to check before anyone else saw it.

If your task is preparing for a coaching conversation, this workflow takes today's method further, with approval gates and source checks. Use general or fictional material only: Workflow: Executive coaching preparation (members’ Studio)

Advanced variation (optional)

Before you run a framed request, ask Claude to critique it first: "Before you answer, tell me what is unclear or missing in this request, and what you would otherwise have to assume." Decide which suggestions to accept and why. You stay the author of the request.

Bring to the next session

One before-and-after pair, with no client names or confidential details, and one statement from a reply that you could not verify. Write both in your Apply reflection below; it will appear at the start of Session 2.