AI Training Hub
View the slide deck →
Facilitator Box Kit · v1.0

Building an
AI Native
Mindset

A complete teaching package for the 3-hour workshop. Designed so any trained facilitator can pick it up and deliver with confidence — opening through capstone, demos through troubleshooting, FAQs through follow-up.

At a Glance
Modules8 + Opening
Time3 hrs
AudienceWorking Pros
ToolsChatGPT / Claude
FormatIn-Person
CapstonePersonal AI System
§ 01 One-Page Snapshot

Course Overview

A facilitator should be able to read this single section and know exactly what they're teaching, to whom, and why.

Session Title
Building an AI Native Mindset
What This Session Is About
Shifting working professionals from casual AI use to systematic collaboration — and from one-off prompts to repeatable AI-native workflows.
Why It Matters
The professionals who learn to think with AI — not just ask it questions — will compound their value over the next decade.
Intended Audience
Working professionals who already use AI casually but aren't yet using it systematically.
Prerequisites
Has used ChatGPT or Claude at least 5 times. Active account they can log into on their laptop. No coding required.
Tools Used
ChatGPT or Claude. Free tier covers Modules 1–5; paid tier helpful for Module 6 demos (file uploads, Custom GPTs / Projects).
Total Time
~3.5 hours, including two 10-minute breaks.
Accessibility
Slides use 24pt minimum. Demos narrated aloud as they happen. Captions enabled if recording. Printed handouts available in 14pt on request.

Learning Outcomes

By the end of the session, participants will be able to:

  1. Shift from one-off prompts to multi-turn collaboration with AI.
  2. Give AI better context, constraints, examples, and source material using the CRAFT framework.
  3. Use structured prompting for more reliable outputs.
  4. Turn their professional judgment into reusable prompts, rubrics, and workflows.
  5. Use AI for research, synthesis, writing, and communication.
  6. Identify three categories of AI risk — hallucination, data exposure, reputation — and apply human-in-the-loop review.
  7. Design one larger AI-native workflow they can continue building after the session (the Personal AI System capstone).

Core Thesis

AI-native professionals don't just use AI to answer questions. They use AI to rethink how work gets done. They look at emails, documents, spreadsheets, meetings, research, and manual data entry and ask: "Can AI summarize, extract, classify, draft, compare, recommend, or automate this?" Course Thesis
§ 02 Minute-by-Minute

Run of Show

If you fall behind, the cues below show what to compress. Never skip Module 3 — it's the workshop's center of gravity.

Time Type Topic / Activity Facilitator Cues
0:00–0:07 Opening
How AI Actually Works
Mental model before any prompting. ML, language models, black box, what AI knows vs. what only you know. Old way vs. AI-native framing.		 5–7 min. Walk Old/AI-Native table aloud with one concrete example per row.
0:07–0:32 Module 1
The AI-Native Mindset
The paradigm shift, not the prompt. Same workflow, you move to the middle.		 25 min. Callback to opening question. Land the "redesign the workflow with AI in it" frame.
0:32–0:57 Module 2
Prompting — CRAFT
Context · Role · Action · Format · Tone. Five letters, five questions, one good prompt.		 25 min. Each letter gets one sentence. Run the before/after demo. Hand out template card.
0:57–1:27 Module 3
Encoding Subject Matter Expertise
Rubrics, checklists, role-specific systems. Build one.		 30 min. ⏰ TIME CHECK at 1:20. Never cut this module.
1:27–1:37 Break
10-minute reset
Coffee · stretch · quick questions		 Reset projector. Pull up Module 4 first slide.
1:37–2:02 Module 4
AI Risks, Hallucinations & Guardrails
Three warnings, one habit. Lawyer case · Samsung case · Sports Illustrated case.		 25 min. Don't moralize. These are warnings, not lectures.
2:02–2:27 Module 5
Research, Synthesis, Writing, Communication
Live transformation of messy notes → polished output.		 25 min. Use a participant's real material if anyone volunteers.
2:27–2:37 Break
10-minute reset
Coffee · stretch · quick questions		 Reset projector. Pull up Module 6 first slide.
2:37–3:02 Module 6
AI Beyond the Chat Window
Where AI is heading next. One live demo of embedded AI (Gmail, Docs, Sheets, etc).		 25 min. Pick ONE tool to show end-to-end. Don't tour.
3:02–3:17 Module 7
Small, Low-Risk AI Projects
5 starter projects. Each participant picks one. Commit before they leave the room.		 15 min. Project cards live on tables.
3:17–3:32 Module 8 + Wrap
Capstone — Your Personal AI System
One workflow, two weeks to v0.1. Workflow mapping. Take-home brief. Final Q&A.		 15 min. Set follow-up cadence: Day 3 nudge, Week 2 share-out.

If You're Behind

At the 1:20 time check (mid-Module 3): cut Module 4's third risk example (Sports Illustrated) to save ~4 min. Run only one Module 5 transformation instead of two to save ~6 min. If behind after the second break: collapse Module 6 to a 15-minute single demo and protect the capstone. Never cut Module 3. Never cut either break.

§ 03 Deck Standards

Slide Deck Guide

The deck is a runway, not a script. One idea per slide. Speaker notes carry the weight.

Design Requirements

One main point per slide
If a slide has more than one headline, split it. Facilitators should never have to choose where to look.
Visual over textual — images, diagrams, key phrases
No paragraphs on slides. If you wrote a sentence, move it to speaker notes. Use icons, real-interface screenshots, and the side-by-side prompt comparison format.
Speaker notes include timing, analogy, and check-in question
Each note block contains: (1) the 1–2 ideas to land, (2) a plain-language version, (3) one analogy, (4) one question to ask the room, (5) common misconception, (6) minutes to spend. Flag skippable slides with ⏰.
Timing cues visible to facilitator only
Use presenter view. Each section title slide shows the cumulative clock (e.g., "00:40 — Module 2 starts here"). Prevents the most common failure: silent rushing.
Flag skippable content explicitly
⏰ TIME CHECK — Skip if running behind. The slides to drop, in order, when time is short: Module 4's third example (Sports Illustrated), Module 5's second transformation variation, Module 6's second tool tour.

Slide Counts by Section

Rough target counts. Adjust to your style.

  1. Opening: 6–8 slides (3 for "How AI Works," 1 for Old/AI-Native table, 2–3 for framing).
  2. Module 1: 4 slides (1 intro, 1 callback to opening question, 2 for the workflow shift diagram).
  3. Module 2: 8 slides (1 intro, 5 for each CRAFT letter, 2 for the template card).
  4. Module 3: 7 slides (1 intro, 2 for rubric building, 2 for role examples, 2 for activity).
  5. Module 4: 5 slides (1 intro, 3 for the failure examples, 1 for human-in-the-loop).
  6. Module 5: 4 slides (1 intro, 1 for transformation patterns, 2 for demo).
  7. Module 6: 5 slides (1 intro, 1 for the landscape, 3 for the live demo).
  8. Module 7: 6 slides (1 intro, 5 project cards).
  9. Module 8 + Wrap: 5 slides (1 capstone intro, 1 worksheet, 2 example capstones, 1 next steps).
§ 04 How To Say It

Facilitator Talking Points

These are how you say it, not just what to say. Use them verbatim if you're new to teaching this material.

Opening · How AI Actually Works (5–7 min)

Before any prompting lessons, learners need a basic mental model. Spend 5–7 minutes here. Goal: by the end, learners understand why hallucinations happen, not just that they happen. This pays off when you reach Module 4.

01 · Machine Learning

It learned, it wasn't programmed.

AI today isn't a giant rulebook. It learned patterns from a huge amount of text — books, websites, code, conversations. Nobody wrote "when someone asks X, say Y." The model figured out X→Y from examples.

02 · Natural Language Processing

The field of teaching machines to read and write.

NLP is the corner of computer science that handles human language. Translation, autocomplete, spam filters, voice assistants — all NLP. ChatGPT and Claude are the most powerful NLP systems ever built.

03 · The Black Box

Even the builders can't fully explain it.

This is the uncomfortable part. The engineers at OpenAI and Anthropic can describe the math, but they can't trace exactly why a model produces a specific sentence in a specific moment. The path from input to output isn't human-readable. That's not a flaw — it's just how these systems work.

04 · What This Means For You

AI predicts plausible next words. It doesn't "know."

The model is a very sophisticated pattern-matcher predicting what should come next. So it confidently produces sentences that sound right but aren't. That's a hallucination. It's not the AI being broken — it's the AI doing exactly what it does.

And critically: AI doesn't have access to anything in your head. It doesn't know your company. It doesn't know your customer. It doesn't know what you decided in Tuesday's meeting. If you don't tell it, it will make a plausible-sounding version up. The bridge to Module 1 — context matters because the AI has none of yours

Opening · Old Way vs. AI-Native

Walk this table aloud. Give one concrete example for each row — your own or one of the examples below. Don't force learners to participate; just narrate.

Old Way AI-Native Way Concrete Example to Use
Search manually Ask AI to search, summarize, and compare Last week I needed to compare three vendors' contracts. Old way: 90 minutes reading them side-by-side. AI-native: paste them in, ask for the differences in plain English, get a 4-bullet summary in 30 seconds.
Rewrite from scratch Give AI context and examples Drafting a quarterly update for the board. Old way: stare at a blank doc for 20 minutes. AI-native: paste in last quarter's update + this quarter's metrics + the board's interests, ask for a first draft in that exact voice.
Do repetitive data entry Use AI to extract and classify 200 customer support tickets need categorizing for the monthly report. Old way: 3 hours of manual tagging. AI-native: paste in the tickets, give the AI your categories, get tagged output in 5 minutes — then spot-check.
Ask one-off prompts Build repeatable workflows Every Monday I prep for my 1:1s. Old way: re-invent the format each week. AI-native: I built a Custom GPT that takes my notes from last week and outputs my agenda. Same workflow, every Monday, 90 seconds.
Keep expertise in your head Encode judgment into prompts and rubrics I've evaluated 200 sales leads in my career. Old way: that expertise lives only in my brain. AI-native: I wrote down my 5 criteria, taught the AI to score new leads with them, and now anyone on my team can run the same evaluation.

Tip: Bring 1–2 of your own examples from the days before the workshop. Personal stories beat hypotheticals. After you walk the table, transition: "Today's job is to move you from the left column to the right. That starts with one shift — treating AI as a collaborator, not a vending machine."

Key Concepts to Emphasize

The four ideas every learner must leave with:

  1. AI is a collaborator, not a vending machine. You don't put a coin in and get an answer. You work with it, give it material, and iterate.
  2. Context is the prompt. The words you type are 20% of the input. What you tell AI about your situation, audience, and constraints — that's the other 80%.
  3. Your expertise is the moat. Anyone can prompt. Only you can encode what makes a good contract, customer email, financial model, or lesson plan. That's the AI-native skill.
  4. Trust, but verify — every time. AI removes execution risk. It increases judgment risk. Always review.

Module 2 · The CRAFT Framework

Five letters. Five questions to answer before you hit enter on any non-trivial prompt.

C
Context
What's the background, situation, audience?
R
Role
Who should the AI act as?
A
Action
What specifically do you want it to do?
F
Format
How should the output be structured?
T
Tone
What voice or register?

How to teach it: Don't list the letters and move on. Walk one example through all five, slowly. Use the Module 1 bad-prompt-vs.-good-prompt example, but explicitly label each letter as you go: "Notice how the bad prompt has zero of these. The good prompt has Context (sales follow-up after discovery), Role (sales manager), Action (write the email), Format (under 150 words, three paragraphs), Tone (warm but professional). That's CRAFT."

Key message: A strong prompt looks like delegation to a smart junior employee. They need the same five things from you that the AI does.

Plain-Language Translations

✗ Don't Say

"LLMs use probabilistic token prediction over a transformer architecture."

✓ Say This Instead

"AI predicts the most likely next word, based on patterns it learned from a lot of text. That's it. That's the trick."

✗ Don't Say

"You need to engineer your prompt with appropriate scaffolding and few-shot exemplars."

✓ Say This Instead

"Treat the AI like a smart new hire. Tell it the role, the goal, the audience, the format. Show it one example. Then let it draft."

✗ Don't Say

"Use retrieval-augmented generation to ground responses in source material."

✓ Say This Instead

"Paste in the document. Tell the AI to only use what's in the document. That's it."

Analogies That Work

Prompting AI is like giving directions to a GPS. The vaguer you are, the worse the route. "Take me to a coffee shop" gets a chain. "Take me to a quiet coffee shop with outlets, under a 10-minute drive, open past 6pm" gets the right one. Use in Module 2 · CRAFT introduction
AI knows everything public — and nothing about you. Treat it like a brilliant new hire on day one. Smart, fast, well-read. Knows nothing about your company, your customer, or what happened in Tuesday's meeting. Tell it those things. Use in Opening · transition to Module 1

Common Misconceptions

"If the AI got it wrong, I need to find a better AI."
No — you almost certainly need to give the AI more context. The model didn't fail; the prompt failed to set it up. Demonstrate this live in Module 1.
"Prompting is just a hack — eventually AI won't need it."
Even when models get smarter, you'll still need to tell them what audience, what constraints, what success looks like. That's not a hack — that's communication. It won't go away.
"AI will replace my expertise."
Inverse. AI commoditizes execution. Your expertise — judgment, taste, knowing what good looks like — becomes more valuable, not less. The people winning are the ones encoding their expertise into the system.
"I shouldn't trust AI for important work."
Correct, with an asterisk. Never trust it blindly. But the alternative isn't "don't use it" — it's "use it with review." Surgeons use scalpels with checklists. Pilots use autopilot with cross-checks. Same principle.
"The AI is just making stuff up — it's unreliable."
It can make stuff up, yes — that's what hallucination is. But "unreliable" frames it wrong. It's predictable in how it goes wrong: specific facts, dates, citations, names. Once you know the failure pattern, you verify the right things and use it with confidence elsewhere.
§ 05 Step-by-Step

Demo Scripts

Every demo has been run twice in advance. Backup screenshots live in the shared drive under /boxkit/demo-backups/.

Demo 01 · Bad Prompt vs. CRAFT Prompt

Module1 — Chatting → Collaborating
ToolChatGPT or Claude (web)
Duration5–7 minutes

Goal

Show — not tell — that context transforms output quality. Visceral, not abstract. Sets up Module 2's CRAFT framework.

Setup

Open a fresh chat window. Project the screen. Have CRAFT prompt pre-written in a notes file so you can paste it cleanly.

Step-by-Step

  1. Tell the room: "I'm going to send the worst version of a real prompt, then we'll fix it together."
  2. Type and send:
    Write me a follow-up email.
  3. Read the output aloud. Ask: "What's wrong with this?" Let the room answer. (Expected: too generic, no context, no audience, no specifics.)
  4. In the same chat, send the CRAFT version:
    CONTEXT: I'm a sales manager at a B2B SaaS company. I just finished a 30-min discovery call with the VP of Operations at a 200-person manufacturing firm. Her main pain: her team spends 6+ hours/week on manual reporting. We discussed our reporting automation product. Next step: propose a 30-min implementation planning call. ROLE: Act as an experienced B2B sales manager. ACTION: Write a concise follow-up email that acknowledges her specific pain, recaps what we discussed, and proposes the planning call with two time options. FORMAT: Under 150 words. Three paragraphs. No subject line yet. TONE: Warm, professional, not salesy.
  5. Read the new output aloud. Pause. Let the difference land.
  6. Ask: "What changed?" Surface the answer: the AI didn't change. What changed was the context, role, action, format, and tone. That's CRAFT — that's what we'll spend Module 2 on.

Expected Output

A 3-paragraph email under 150 words. First paragraph acknowledges the manual-reporting pain. Second paragraph recaps the product fit. Third paragraph proposes two specific meeting time options with a warm close.

Common Failure Modes

(a) AI tool is slow or queued — narrate the wait, don't panic. (b) Output is unexpectedly bad — lean in: "Even good prompts can produce weird first drafts. Watch how I'd iterate." Send one follow-up clarification. (c) Tool crashes — backup plan below.

Backup Plan

If tool is down: switch to Slide 8, which has pre-captured screenshots of both outputs side-by-side. Narrate the same beats. Total recovery time: under 30 seconds.

Simpler Version

If you're short on time: skip the CRAFT labels in the prompt and just send a denser prompt. Don't explain CRAFT until Module 2. Saves 2 minutes, but you lose the explicit bridge.

Demo 02 · Building a Rubric Live

Module3 — Encoding Expertise
ToolChatGPT or Claude (web)
Duration8–10 minutes

Goal

Demonstrate the leap from "asking AI to evaluate" to "giving AI your evaluation criteria." This is the workshop's pivotal moment.

Setup

Whiteboard ready. Markers in 2 colors. Fresh chat window open. Anonymized sample lead description ready to paste.

Step-by-Step

  1. Ask the room: "Sales folks — what makes a lead worth pursuing?" Collect 4–5 criteria on the whiteboard. (Expected: budget, authority, timing, fit, signal of pain.)
  2. Tell them: "We just built a rubric. Now we'll teach the AI to use it."
  3. Send this prompt:
    CONTEXT: I'm a senior sales rep evaluating an inbound lead. I want a consistent scoring approach my whole team can use. ROLE: Act as a senior sales qualifier. ACTION: Use the rubric below to score the lead I describe. For each criterion, give a 1–5 score and a one-line justification. Then a total. Then recommend: pursue / nurture / disqualify. RUBRIC: — Budget signal (is there money?) — Authority (decision-maker engaged?) — Timing (urgent vs. exploratory?) — Fit (do we serve their industry/size?) — Pain signal (real problem stated?) FORMAT: Markdown table for the scoring; recommendation in bold below. TONE: Direct, analytical. LEAD: [paste anonymized lead description]
  4. Read the scored output aloud. Let learners see the AI applying their criteria.
  5. Land the point: "You just turned a decade of sales judgment into a reusable system. Anyone on your team can run this now. That's encoding expertise."

Expected Output

A 5-row markdown table with scores, justifications, a total, and a bolded recommendation. The output should feel like the facilitator's judgment, not the AI's opinion.

Common Failure Modes

(a) Room is quiet, no one offers criteria — use the prepared customer-success health score rubric from Slide 22. Same demo, different domain. (b) AI scores feel "off" — that's a feature, not a bug. Say: "What's missing from the rubric? That's the work — refining until the AI's judgment matches yours."

Backup Plan

If tool fails: same screenshots in /boxkit/demo-backups/demo02-rubric.png. Narrate the same beats with the static image.

Simpler Version

If running short: use the pre-built lead-qualification rubric on Slide 22 directly. Skip the whiteboard collection step. Saves 4 minutes, loses some engagement.

Demo 03 · Embedded AI Live (Gmail Triage)

Module6 — Beyond the Chat
ToolGmail + Claude/ChatGPT side-panel
Duration6–8 minutes

Goal

Show that AI's value compounds when it lives in your tools — not in a separate browser tab.

Setup

Sample Gmail inbox with 8 staged emails of varying urgency (DO NOT use a real personal inbox). AI assistant feature pre-enabled. Browser zoomed to 125% so the room can read text.

Step-by-Step

  1. Tell the room: "I want AI to triage this inbox for me — flag urgent, summarize the rest, draft replies for the easy ones."
  2. Open the AI side panel in Gmail. Type the triage prompt (paste from kit):
    Look at my last 8 unread emails. Tell me: which 1–2 are urgent (and why), what the rest are about in one line each, and draft a 2-sentence reply for any that are simple FYI confirmations.
  3. Let the output stream. Read key parts aloud. Don't over-narrate; let the speed and integration speak for itself.
  4. Pick one drafted reply. Click it. Show the human-in-the-loop edit — change one phrase before "sending."
  5. Land the point: "8 emails handled in 90 seconds, with judgment intact. The AI lives where the work happens. That's the frontier."

Expected Output

A structured response with 2 urgent flags, 6 one-line summaries, and 2–3 short reply drafts. Visibly faster than the room would have done manually.

Common Failure Modes

(a) AI feature not showing in side panel — check feature is enabled in tool settings before the workshop. (b) Output is empty — re-prompt with more specificity, treat as a teachable moment. (c) Tool is slow — narrate the wait, use the time to ask the room: "What's your own email backlog look like right now?"

Backup Plan

If the live demo fails completely: switch to the screen-recording at /boxkit/demo-backups/demo03-gmail.mp4. Play with narration. Acknowledge the irony briefly ("The frontier is great when the WiFi works") and move on.

Simpler Version

If you don't have access to a Gmail account with an AI side panel: open Claude or ChatGPT, paste in the same 8 emails as text, run the same prompt. Loses the "embedded in tools" point but keeps the workflow point.

§ 06 From Listening to Doing

Practice Activities

One activity per module. Don't let participants stay passive past the 30-minute mark.

Module 1 · 8 minutes

Map Your Old Way → AI-Native Way

Objective: Identify one workflow you currently own end-to-end, and mark which steps could shift to AI — putting you in the middle as the judgment layer.

Instructions

  1. Pick a workflow you do at least weekly. Write the steps in order, left to right.
  2. For each step, mark it YOU (judgment, taste, decision) or AI (pull, clean, format, draft, classify).
  3. Find the one step that's currently YOU but could be AI. Circle it.
  4. Share with one neighbor: which step did you circle? What's stopping you from making the switch on Monday?

Expected Output

Every learner has at least one concrete step they can hand off to AI this week. The "aha" is recognizing how many clerical steps they were doing out of habit, not necessity.

Debrief Questions

"What surprised you about how many steps could shift? Was anyone tempted to mark a judgment step as AI? Where's the line for you?"

Differentiation

Stuck? Use one of the laminated "starter scenario" cards on each table (5 pre-written workflows across sales/ops/marketing/HR/exec).

Module 2 · 10 minutes

CRAFT-ify One of Your Own Prompts

Objective: Take a vague prompt you sent this week and rewrite it with all five CRAFT fields. Save it for Monday.

Instructions

  1. Pull up your AI tool's chat history. Find any prompt you sent this week. Pick a vague one.
  2. On the worksheet, fill in all five fields: Context · Role · Action · Format · Tone.
  3. Combine them into one prompt. Run it. Compare to the original output.
  4. Save the new prompt — notes app, doc, or a Custom GPT. Share the biggest delta with one neighbor.

Expected Output

Every learner has one saved, working CRAFT prompt they can use on Monday. That's the bar.

Debrief Questions

"Which letter was hardest to fill in? Most people skip Tone. Did anyone find that adding Tone changed the output more than expected?"

Module 3 · 12 minutes

Encode One Piece of Your Expertise

Objective: Build one rubric, checklist, or template from your job that AI can run.

Instructions

  1. Think of one recurring evaluation you do at work. "Is this a good [email / PR / contract / candidate / vendor]?"
  2. Write 3–6 criteria you use — even if you've never written them down before. That's the point.
  3. Wrap them in a CRAFT prompt. (Context: you're evaluating Xs. Role: senior reviewer. Action: score using the rubric. Format: table. Tone: direct.)
  4. Test it. Paste in a real (anonymized) example. See if the AI's evaluation matches yours.
  5. Refine. If the scoring feels off, your rubric is missing something. Add it.

Expected Output

Each learner leaves with a working rubric they can save, share with their team, and reuse weekly.

Differentiation

Stuck? Hand them the "ten common professional rubrics" card (email triage, meeting prep, weekly review, vendor scoring, candidate screen, etc.). Note: the Module 3 example sets (sales, product, ops) should be swapped to match the actual roles in the room. Alternate sets available for legal, finance, HR.

Challenge Extension

Turn the rubric into a Custom GPT or Claude Project. Sets up the take-home capstone.

Module 4 · 5 minutes

Risk-Check Your Own Prompt

Objective: Apply the three risks (hallucination, data exposure, reputation) to a prompt you've built today.

Instructions

  1. Pick one prompt or workflow you built in Modules 1–3.
  2. Answer three questions on the worksheet: (a) Where could it hallucinate? (b) What data should never go in? (c) What's the cost if the output goes out unedited?
  3. Identify one place to add a human-in-the-loop checkpoint.

Expected Output

Each learner has annotated one of their workflows with a review checkpoint.

Debrief Questions

"Where in your workflow does a mistake stay invisible? That's the highest-stakes review point."

Module 5 · 8 minutes

Messy Notes → Polished Output

Objective: Use AI to transform raw material into something you'd actually send.

Instructions

  1. Open something messy on your laptop: meeting notes, a long email thread, a brain-dump doc.
  2. Write a CRAFT prompt to turn it into: a summary, a list of decisions, action items, or a recap email — your choice.
  3. Run it. Edit the output to be 90% there. Note: you're not writing from scratch.

Expected Output

One polished artifact per learner, going from messy to shippable in under 8 minutes.

Differentiation

If they don't have messy material handy: provide a sample transcript or notes file on each table.

Module 7 · 10 minutes

Pick Your Starter Project

Objective: Each learner commits to ONE low-risk starter project to build this week.

Instructions

  1. Read the 5 project cards on your worksheet (Gmail triage / Meeting notes / Document digest / Personal prompt library / Spreadsheet cleanup).
  2. Pick the one that maps to a real pain you have this week.
  3. On the worksheet, write the first CRAFT prompt you'd use.
  4. Share with one neighbor — they critique your scoping.

Expected Output

Every learner names a project and drafts at least one working prompt for it before they leave the module.

Differentiation

For learners stuck: hand them the "default project" card pre-scoped for their role.

§ 07 Take-Home

Capstone · Personal AI System

The take-home project that turns one workshop into a lasting practice.

The Brief

Pick one recurring workflow you own. Redesign it as an AI-native workflow. Ship a working v0.1 within two weeks. Capstone Charter

In-Session Worksheet Structure (15 min)

During the wrap, learners fill in:

  1. Manual process today. What happens step by step right now?
  2. Inputs. What information / files / signals start the process?
  3. Bottlenecks. Where does it slow down or stall?
  4. Repetitive decisions. What judgment calls do you make over and over the same way?
  5. Human judgment points. Where must a human stay in the loop?
  6. AI-assisted steps. Which steps can AI summarize, extract, classify, draft, compare, or recommend?
  7. Tools needed. What's the minimum tech to make this work?
  8. First prototype. What's the smallest version you can build this week?

Example Capstones

Adapt these to your cohort's roles. Use as conversation starters during the activity.

Sales
AI Account Executive Assistant
Research the account · identify buying triggers · draft outreach · prep discovery questions · summarize calls · update CRM. End-to-end for one rep's daily flow.
Product
AI Product Discovery Assistant
Analyze customer feedback · cluster pain points · score opportunities · draft a PRD · build the stakeholder brief.
Operations
AI Process Automation Map
Pick a repetitive process · map every step · identify extract/classify/draft handoffs · design the human review loop · ship v0.1.
Executive
AI Chief of Staff Briefing
Summarize email, calendar, docs, priorities · identify open decisions · draft responses · track open loops across a week.

Follow-Up Cadence

  1. Day 3: Short email — "What did you pick? What's your first prompt?"
  2. Week 2: Optional 30-min virtual share-out — 5 learners demo their v0.1.
  3. Week 6: Retro — what worked, what didn't, what they shipped.
§ 08 Temperature Reads

Check for Understanding

Use these throughout — not at the end. The cost of a confused learner at minute 30 is a lost learner at minute 90.

Quick Visual Checks

  • "Thumbs up if that prompt made sense before I explained it."
  • "Hold up 1–5 fingers — how confident are you that you could write a CRAFT prompt for your own work right now?"
  • "Raise a hand if you've ever sent a prompt as vague as the first one I showed." (Almost everyone will. That's the point.)
  • "Look at your neighbor's worksheet. Thumbs up if their rubric makes sense to you."

Mini Challenges

  • "In the next 60 seconds, add one more letter from CRAFT to your prompt. Ready, go."
  • "Take the bad prompt on screen. Tell your neighbor the first thing you'd add."
  • "Predict: if I add a Tone line, what will change in the output?"

Reflection Questions

Why do you think the CRAFT prompt worked better? What changed — the model, or what we gave it? What surprised you? Use after Demo 01 · Module 1/2 bridge

When Learners Are Confused

If a learner can't articulate what changed in the demo
Walk them backward. "Forget the AI for a second. If I asked a junior teammate to 'write me a follow-up email,' what would they ask me before writing? That list of questions is CRAFT."
If a learner says "this won't work for my job"
Take it seriously. Ask them to describe one recurring evaluation, decision, or piece of writing they do. There will be a rubric or template inside it. Sit with them for two minutes if needed.
If the room is silent after a prompt
Don't repeat the question — that signals it was bad. Reframe: "Let me put it this way…" or call on a table specifically: "Table 3, talk to each other for 30 seconds, then I'll come back."
§ 09 Keep the Energy

Engagement Strategies

A dead room kills learning. Reset energy every 8–10 minutes, especially after lectures.

Think — Pair — Share
"Take 60 seconds to think about this on your own. Then turn to your neighbor and tell them your answer. We'll hear from two tables." Use after any conceptual block longer than 4 minutes.
Low-Pressure Entry
"Take 30 seconds and just try it — it doesn't need to be good." Or: "Talk to your table first. We'll hear from groups, not individuals." Removes the spotlight.
Calling on Learners
Casual, never aggressive. "Any volunteers?" → if silence, "What do you think, Table 2 — just throw something out." Cold-call tables, not individuals.
Prediction Questions
Ask before showing the answer. "What do you think will happen if I add a Tone line to this prompt?" Then reveal. Engagement triples when the learner has skin in the game.
Energy Reset
Shift activity every 8–10 minutes. Demo → discussion → practice → quick poll. Never lecture for more than 12 minutes without a beat change.
The "Real Example" Trump Card
If the room is theoretical, ask: "Who in here actually sent an AI prompt this morning?" Get one volunteer. Run their prompt live. Real beats abstract every time.
Circulate During Activities — Never Sit
During Modules 3, 7, and 8 (the hands-on stretches), walk slow loops. Stop where you see someone staring. Ask: "What did you pick?" Listen for vagueness. Push for specificity.
§ 10 Contingency Plans

What Could Go Wrong

Every scenario below has happened in a real session of this workshop. The recovery is your job — the plan is ours.

AI tool crashes mid-demo
Switch immediately to backup screenshots in /boxkit/demo-backups/. Don't try to debug live — that kills 5 minutes. Narrate the screenshots with the same beats. Move on within 30 seconds.
Wi-Fi or projector fails
Pivot to printed worksheets. Use the whiteboard to write CRAFT and the rubric example. Module 3 runs entirely offline if needed — it's the most worksheet-friendly module.
Running behind by 10+ minutes
If you catch it at the 1:20 time check (mid-Module 3): cut Module 4's third risk example (Sports Illustrated) ~4 min, run only one Module 5 transformation instead of two ~6 min. If you catch it after the second break: collapse Module 6 to a single 15-min demo and protect the capstone. Never cut Module 3 or either break.
Finishing early by 15+ minutes
Open the floor to "show me a prompt you've sent." Workshop them live with the room. Most engaging way to fill time and reinforces CRAFT.
Learners visibly lost
Stop. Don't push through. Ask: "What's the one thing that just stopped making sense?" Re-explain that one thing with a different analogy. The rest of the session depends on this.
Quiet room, no one answering
Switch to table-level. "Table 1, give me one answer." Pairs and small groups always outperform full-room cold-calls. If still quiet, name it: "We're warming up. That's normal."
A learner is monopolizing
Validate, then redirect. "Great point — let's hear from someone else who's been quieter." Don't shame. If it continues, talk to them at the break privately.
Hostile or skeptical learner ("AI is hype")
Don't argue. "Fair — and here's a real workflow from a peer in your role that saved them 6 hours/week. Try it for a week, judge the data." Move on.
Demo produces a bad/weird output
Lean in. "Perfect — this is exactly why we don't ship anything without review. What's wrong with this output?" Turn the failure into Module 4 (risks) early.
Mixed cohort — some way ahead, some way behind
Pair fast learners as table mentors. Hand stuck learners the "starter scenario" cards. Use challenge extensions to keep the fast ones busy without leaving the slow ones behind.
A learner wants to paste real confidential data
Stop them gently. "Hold on — let's not put that in. Use the anonymized version on the worksheet, or describe the situation without specifics. Module 4 covers why."
§ 11 Anticipated Questions

FAQ & Q&A

20 questions facilitators will hear. Memorize the first six.

Basic / Beginner Questions
"Which AI tool should I be using — ChatGPT, Claude, Gemini, Copilot?"
For general work: all four are strong. For long documents and nuance: Claude. For Microsoft ecosystem: Copilot. For Google ecosystem: Gemini. For broadest plugins: ChatGPT. Pick one, use it daily for a month, then evaluate. Switching tools is the procrastination of AI adoption.
"Will the AI use my prompts to train future models?"
Depends on tool and account. ChatGPT free: yes by default (toggleable). ChatGPT Team/Enterprise: no. Claude consumer: no by default. Claude API: no. When in doubt, check settings, and never paste confidential client data into a personal free account.
"What if I'm worried about confidentiality?"
Three-tier rule: (1) Public info — fine anywhere. (2) Internal but non-sensitive — enterprise tools only. (3) Confidential or regulated — never in a third-party tool without legal sign-off. Module 4 covers this.
"How long should a CRAFT prompt be?"
As long as it needs to be. Most professionals undershoot by 80%. A good prompt for a real work task is 100–300 words. If you're typing one sentence, you're under-investing.
"Why does the AI sound the same as everyone else's AI?"
Because most people skip the Tone letter in CRAFT. Tell it your voice. Give it three examples of how you write. Tell it what you'd never say. The "AI tone" is the absence of instruction, not a fixed feature.
"Can the AI access the internet?"
Sometimes. ChatGPT, Claude, and Gemini all have web-search modes. Default chats may or may not — check for a globe/search icon. For research tasks, explicitly enable web search.
"Should I tell people I used AI?"
Depends on context. For drafts you reviewed: usually no — same as you wouldn't disclose spell-check. For published or attributed content: increasingly yes. For internal team norms: ask your manager. Default to transparency when uncertain.
"What's a Custom GPT or Claude Project?"
A saved AI environment with persistent instructions and uploaded files. Once you find a prompt you use weekly, turn it into a Project. That's the upgrade path. Module 6 demos this.
"What if the AI gives a bad answer?"
First, treat it as a prompting problem — was your CRAFT prompt complete? 80% of bad answers come from incomplete prompts. If the prompt was solid, iterate: "That missed X. Try again with…" Multi-turn beats one-shot every time.
"What data should I never put into AI?"
Personally identifiable customer data (names, SSNs, financial details), confidential client work without permission, regulated data (HIPAA, FERPA, financial records depending on your industry), source code with proprietary algorithms, and unreleased financial info. Anonymize first, or use enterprise tools.
Advanced / Challenge Questions
"How do I know when the AI is hallucinating?"
Default assumption: any specific claim — date, statistic, citation, quote, name — might be wrong. Verify everything specific. The more confident the AI sounds, the more carefully check. We covered why in the Opening: it predicts plausible next words, not facts.
"What about agents — should I use those instead?"
Not yet, for most professional work. Agents (AI that takes actions in tools) are improving fast but failure modes are still expensive. Build the prompt and workflow skills first. Agent skill is downstream.
"How does this differ from RAG / fine-tuning / custom models?"
Those are engineering paths — different skill, different cost. For 95% of professionals, CRAFT prompts + uploaded documents + Projects are enough. We're teaching applied use, not ML engineering.
"Can I trust AI for legal / medical / financial work?"
Use it to draft, summarize, research, or compare. Never to decide. The professional in the room (the lawyer, doctor, advisor) is the decision-maker. AI is the associate, not the partner.
"What if my company doesn't have an approved AI tool?"
First, ask IT or InfoSec — most companies are further along than employees realize. If truly no tool: use a personal account for non-confidential work only, and start the conversation with your team. Shadow AI use is the riskiest path.
"Will AI write better prompts than I will eventually?"
Sort of — already happens. You can ask AI to improve your prompt. But you still need to know what good output looks like, what audience, what to include and exclude. That's the AI-native skill: judgment, not typing.
"How do I get my team / boss to take this seriously?"
Show, don't sell. Do the work in half the time using AI. Bring your CRAFT prompts to your next 1:1. Don't pitch "AI strategy" — demonstrate "I shipped this twice as fast and here's how."
"How do I keep this practical instead of theoretical?"
Pick ONE workflow this week. Ship v0.1 by Friday. The capstone is built for this. Theory without shipping is procrastination dressed up.
"What if participants are way more advanced than the material?"
(Facilitator question.) Pair them as table mentors. Use the challenge extensions in each activity. Their job becomes helping the rest of the table — which deepens their own thinking.
"That's a great question — but I don't know."
(A reminder, not a learner question.) It's okay to say: "I don't know — I'll follow up." Facilitators don't need to know everything. Note the question, follow up by email within 48 hours. Builds more trust than guessing.
§ 12 Plain-Language Definitions

Glossary

Use these definitions when learners ask. Resist the urge to over-explain.

AI-Native
A way of working where AI is woven into how you do tasks — not a tool you occasionally consult. The shift from "I'll ask AI about this" to "I'll design this workflow with AI in it."
Machine Learning (ML)
The branch of AI where systems learn patterns from data rather than being programmed with explicit rules. ChatGPT and Claude are products of machine learning at enormous scale.
Natural Language Processing (NLP)
The field of getting computers to read, understand, and write human language. Translation, autocomplete, spam filters, voice assistants — all NLP. Modern AI chatbots are the most powerful NLP systems ever built.
LLM (Large Language Model)
The kind of AI behind ChatGPT, Claude, and Gemini. Predicts the most likely next word based on patterns learned from a huge amount of text. The "intelligence" is pattern completion at scale.
The Black Box
The fact that even the engineers who build modern AI models can't fully explain why a specific output happens in a specific moment. The math is understood; the path from input to output isn't human-readable.
Prompt
What you type into the AI. The instructions, context, and material you give it before it responds.
CRAFT
The five-part prompt framework taught in this workshop: Context, Role, Action, Format, Tone. Answer all five before sending any non-trivial prompt.
Context
The background information that helps the AI understand your situation. Audience, constraints, goals, source documents, examples. The most underused lever in prompting.
Multi-turn
A back-and-forth conversation with AI rather than a single message. The AI remembers earlier messages in the same chat and uses them as context.
Hallucination
When AI generates content that sounds confident but is factually wrong. Common with dates, statistics, citations, and obscure names. Default assumption: verify everything specific.
Rubric
A set of evaluation criteria. The structured way you (or your team) decide whether something is good. When you give the AI a rubric, you're encoding your judgment so it can apply it consistently.
Custom GPT / Claude Project
A saved AI environment with persistent instructions and uploaded files. Like having a coworker who remembers your team's playbook every time you ask them something.
Human-in-the-Loop
The principle that AI drafts, humans decide. Especially for anything with consequences — contracts, customer communication, financial calls, hiring.
Embedded AI
AI that lives inside the tools you already use — Gmail, Excel, Salesforce, Slack — rather than in a separate chat window. The frontier this workshop is pointing at.
Agent
An AI that doesn't just respond — it takes actions in tools on your behalf. Sends emails, updates records, books meetings. Powerful and brittle; introduce slowly.
Token
The unit AI uses to read and write — roughly 0.75 of a word in English. You'll see this in pricing and length limits. Mostly not something to think about day-to-day.
§ 13 Pre & Post Flight

Before & After

Click each item as you complete it. Saved locally in your browser session.

Before the Session

  • Review slides end-to-end (60 min)
  • Practice all three demos twice
  • Test all tools and logins
  • Open all links in advance tabs
  • Check internet, audio, projector
  • Print worksheets (1 per learner + 5 spare)
  • Prepare 2 personal examples for Old/AI-Native table
  • Capture fresh backup screenshots
  • Confirm Module 3 examples match cohort roles
  • Set timer for module transitions
  • Review "What Could Go Wrong" section
  • Bring whiteboard markers (3 colors)
  • Eat. Hydrate. You'll be talking for 3 hours.

After the Session

  • Send recap email within 24 hours
  • Share CRAFT card and prompt templates
  • Send Day-3 capstone nudge email
  • Schedule Week-2 share-out
  • Send post-session feedback form
  • Complete facilitator reflection (below)
  • Log any tech issues for ops team
  • Note unclear content for next revision
  • Update FAQ with new questions surfaced
  • Flag standout learners for case-study use
  • Return materials to shared kit
  • Take 30 minutes to decompress

Facilitator Reflection

Complete after every session. Five minutes. Honest.

  1. What worked well?
  2. Where did learners struggle most?
  3. How specific were the capstone workflows — vague, mixed, or specific?
  4. Which capstone surprised you — worth featuring next time?
  5. Timing adjustments needed?
  6. One quote, observation, or question worth saving for the FAQ?