1. Less busywork.
More human work.
1:00–1:03. Welcome everyone. Say: Last time you saw a lot in an hour. Today we slow down, try things together, and leave with something useful. Ask everyone to open ChatGPT and this workshop site on their laptop. All exercises use fictional information. Questions are welcome throughout. Ask for one task each person wishes took less time; take three quick answers and keep the others for the clinic.

2. Watch it. Try it. Make it yours.
1:03–1:05. Breaks are 2:00–2:10 and 3:40–3:50. Show /workshop/ now: prompts, sample cases and instructions live there so nobody has to retype from the projector. Success today means one reusable instruction set, one reviewed result and one task each person wants to keep using. This is a workshop, not a test. If a live tool fails, keep the lesson moving with the written exercise.

3. Start with what is possible
1:05–1:06. Acknowledge the month since the first session. No one needs to remember the terminology. We will separate asking for help, giving the assistant tools, and running work on a schedule.

4. I finally described
what I wanted.
1:06–1:11. This is TJ's personal story, not a claim about every user's experience. Explain: I tried setting up Notion about once a year for four or five years and always found it too complicated. This time I dictated the outcome I wanted and worked with Claude to build it. Show one prepared, non-sensitive Notion view and one useful action, not a tour of your business. Ask: What software have you avoided because setting it up feels like a job? Fallback: tell the before/after story and describe one concrete workflow; do not spend time fixing the demo.

5. Ask. Act. Automate.
1:11–1:15. Use one example across all three: draft an interview invitation; create an event through a connected calendar; prepare a recurring morning brief. Having a conversation does not automatically connect email, calendars or other chats. Ask the room which of the three they have tried. Keep definitions to a sentence each.

6. It needs the brief
you would give a colleague.
1:15–1:19. Explain simply: the assistant produces answers from the information and tools available to it. It can sound certain when wrong. More words are not automatically better; relevant facts and examples are. Show the two requests without typing a long prompt. Ask: What would a new colleague need before writing this for you?

7. Let AI interview you.
1:19–1:22. Run this prompt live on a fictional new-hire welcome email. Answer two questions aloud. Show how the assistant helps discover missing context. Tell it to stop interviewing once the audience, desired result and source facts are clear. Fallback: ask these questions yourself and put the answers in a normal chat.

8. Bring one annoying task.
1:22–1:27. Everyone copies the interview prompt from /workshop/ or uses the previous slide. Allow three minutes to answer, one minute to ask: Summarize the task and propose a small first step, and one minute for a volunteer to share. Completion check: participant can point to a correct task summary and one useful next step. If it guesses, say: That is not what I meant; ask me about the missing detail. Fallback: pair with a neighbor if login or network is unavailable.

9. Dictation and conversation
solve different problems.
1:27–1:31. Demonstrate dictating a short fictional onboarding request through Wispr Flow into a text box; read the resulting text before sending. Then explain that ChatGPT Voice is a conversation, not simply another keyboard. Avoid installation detours. If Wispr is unavailable, use ordinary dictation or type the same rough brief and keep the contrast.

10. Practice the conversation
before the conversation.
1:31–1:37. Turn speaker volume up and open ChatGPT Voice. State clearly that this is a fictional practice scenario, not an actual employee case. Give the fictional change: a meeting moved from morning to afternoon. Speak for two exchanges, then ask for feedback. Ask the room what they would practice. Completion check: name one thing the coaching helped you notice. Fallback: run the role-play in text and have a volunteer read the employee's lines aloud. Do not suggest it substitutes for HR judgment.

11. Ask for help
where the work happens.
1:37–1:41. Open the browser setup already installed on TJ's machine. Explain the value without promising that every website works: understand this page, locate a setting, draft entries, check the result. Ask: What software do you hate using? Take two answers. Explicitly tell the assistant which page to use and what it should stop before doing.

12. Explain it. Prepare it. Check it.
1:41–1:47. Use a harmless practice form, never an actual candidate or employee record. Prompt: Explain this page in plain English. Then prepare the form using these fictional details. Do not submit it. Tell me what is missing. Narrate the instruction, observe the action and inspect the filled fields. Invite the room to spot a missing detail. Fallback after one failed attempt: paste a screenshot or the field labels into chat and have AI explain them and draft the entries. Be explicit that the fallback demonstrates understanding, not browser control.

13. Three places.
One conversation-driven workflow.
1:47–1:52. Show the actual three-panel workspace: chats on the left, active conversation, in-app browser. Open a non-sensitive manager chat or a prepared demo. Frame the future-of-work point as your opinion, not a forecast or promise about every product. Show a simple question about a public page and point to its answer on the page. Fallback: describe the three visible areas using this slide rather than troubleshooting the layout.

14. One manager.
Focused workers.
1:52–1:57. Show one of TJ's actual pinned product manager chats and its completed read-only worker example if visible. Explain: this is how TJ's Codex setup is configured; a normal chat with instructions is not automatically a background team. For an HR audience the transferable idea is a reusable home for a recurring responsibility. Avoid opening customer data or launching a product build in the room. Fallback: walk through a fictional onboarding manager delegating a checklist draft.

15. What would you try
on your own work?
1:57–2:00. Ask two people to name a specific task and the first instruction they would give. Answer one short question. Completion check for the first hour: each person has tried the interview prompt and can name one application. Park deeper questions for the build clinic.

16. Back at 2:10.
2:00–2:10. First ten-minute break. Leave the workshop companion open on the trainer laptop. Help with a small number of access issues if needed, but protect the break and restart at 2:10.

17. Build your reusable workspace
2:10–2:13. Introduce the next forty minutes: everyone will leave with an HR instruction block and a tested draft. Work in a ChatGPT project if available, or one ordinary chat using the same instructions. The exercise works either way.

18. Five parts. Plain language.
2:13–2:18. Walk through this fictional example. Point back to the prior Zoom framework so it feels familiar. Explain that guardrails are practical working instructions, not magic protection. Ask what an onboarding message must never invent: start date, reporting location, pay, policy, contact details. The exercise uses fictional logistics and makes unknowns visible.

19. Write your assistant’s brief.
2:18–2:28. Copy the workspace/context exercise from /workshop/. Prompt: Help me write reusable instructions for an HR assistant. Interview me one question at a time about my role, audience, recurring tasks, writing preferences and facts you should never guess. Then produce an instruction block I can save. Use fictional or non-sensitive examples. Allow six minutes interviewing, two minutes reviewing and two minutes revising. Completion check: participant has a concise instruction block with an explicit missing-information rule. Fallback: fill the five-part brief manually and paste it into a regular chat.

20. Give the work a home.
2:28–2:35. Show where the reusable instructions go in the actual interface available today. Participants follow or paste the block into a dedicated chat. Prompt: Using these instructions, draft a welcome message from the fictional facts in the workshop. Show missing details separately. Avoid implying that a project automatically shares everything with teammates or other chats. Completion check: a saved home for the instructions and a first draft using them. Fallback: save the instruction block in a local note and use an ordinary chat.

21. The second instruction
is often the useful one.
2:35–2:43. Ask participants to make one targeted revision to their welcome message. Prompt: Check this draft against the supplied facts. List anything you assumed. Then rewrite it in a warm, direct tone under 90 words with one clear next step. Allow four minutes revision, two minutes source comparison and two minutes pairing. Completion check: the second version is visibly better and any unsupported detail has been removed or labeled. Fallback: revise one sentence together on the projector.

22. Would you actually use it?
2:43–2:50. Give pairs four minutes, then ask two people to share one improvement. Do not ask people to expose real work or private records. Completion check: each person can name one useful instruction to keep. Transition: now we will apply the same approach to Tony's assessment idea, with a much more precise definition of good.

23. Build the assessment assistant
2:50–2:54. Explain the request as Tony described it: summarize strengths, gaps, concerns and a suggested score from submitted assessments. Today we build the review workflow with fictional samples. We are not asking AI to decide who to hire. The useful outcome is a consistent, inspectable draft for the HR reviewer.

24. A score needs
something behind it.
2:54–2:58. Contrast a bare 9/10 with a review someone can check. Explain that scores are only as meaningful as the agreed task and rubric. No personality guesses, protected-trait inference or automatic hire/reject recommendation. Ask the room which observations would actually help them review an assessment. Keep the output to one screen or page per case.

25. Agree on the measuring stick.
2:58–3:04. Use the companion's fictional work-sample case and anchors. Discuss each criterion with the room. Practice scale: 0 = explicitly unsuitable approach supported by the text; 1 = partial approach with material gaps; 2 = workable approach with minor gaps; 3 = clear, specific, complete approach. Absence of enough evidence is unscored, never automatically zero. A /10 summary is provisional and only shown when all four are assessed: total /12 ×10. Do not normalize away missing criteria. This is a teaching rubric, not a validated employment assessment.

26. Same task.
Two fictional responses.
3:04–3:10. Open the assessment materials in /workshop/. Read the scenario and both responses aloud briefly. Ask participants to identify one evidence passage for a criterion before AI does anything. Then paste the rubric and Candidate A into the assistant. Completion check: everyone can locate the source response and the rubric being used. Fallback: copy the case from the companion into one shared chat and work as a group.

27. Build the first review.
3:10–3:22. Use the full assessment prompt in /workshop/. Core instruction: Apply only the supplied job-task rubric to this fictional response. For each criterion quote supporting evidence, suggest a score using the anchors and explain it. Mark insufficient evidence unscored. Do not infer personality or protected traits. Give strengths, gaps and follow-up questions. Do not recommend hire or reject. Show a provisional /10 only if all four criteria can be scored, using total /12 ×10. Allow eight minutes build and four minutes individual checking. Completion check: an evidence-linked review with no unexplained score. Fallback: complete one criterion together in a table.

28. Find one thing
you would challenge.
3:22–3:32. Pair participants. Use Candidate B and identical instructions. Prompt: Audit your review against the rubric and source. Identify unsupported statements and inconsistent use of anchors. Correct them without inventing evidence. Allow five minutes comparing, three minutes improving and two minutes shareback. Completion check: each pair names one challenged observation or explains why a specific score is supported. Different model outputs are a reason to inspect evidence, not proof one candidate deserves a different result. Fallback: TJ deliberately removes a supporting quote from a sample output and asks the room what is missing.

29. One shared recipe.
Separate candidate reviews.
3:32–3:40. Answer Tony's team question directly: start by distributing the same reviewed instruction block and rubric; each employee can reuse it in their approved tool. A shared project or assistant can package that recipe where available, but uploading a case must respect the organization's access rules. Today save the fictional version. Ask two people to explain the review output and one remaining concern. Completion check: a reusable master plus a clear rule that the person, not the assistant, owns the decision. Fallback: copy the master instructions into a shared approved document; do not pretend a shared product has been configured.

30. Back at 3:50.
3:40–3:50. Second ten-minute break. Preserve the final clinic time. During the break open the fictional scheduling exercise and the morning brief prompt so transitions are fast.

31. Make the useful work repeatable
3:50–3:52. Thirty-five minutes for three practical examples. Scheduling coordinates inputs, a morning brief repeats useful research, and an ideas assistant applies saved instructions to something new. Avoid a tour of settings or subscriptions.

32. Proposing a time
is not booking a meeting.
3:52–3:57. Explain that coordination is straightforward once the required availability and access exist. The assistant cannot inspect calendars it has not been given. A Zoom link is a separate thing to verify, not text to invent. Demo with fictional availability first. If showing an actual connected calendar, use a private test event without external attendees and only if already prepared; otherwise keep it a planning exercise.

33. Find a time.
Make the next step obvious.
3:57–4:05. Use the scheduling sample at /workshop/. Prompt: Using only these fictional availability windows, propose two times for a 30-minute meeting. State the time zone. Show how each time fits everyone. Draft an invitation. Do not claim anything is booked or invent a Zoom link. If fewer than two options exist, say so. Six minutes doing and two minutes reviewing. Completion check: at least one valid overlap, with the draft clearly labeled unsent. Fallback: draw the time-window overlap on the board and ask AI only to draft the invitation.

34. A trigger. A job.
Somewhere you will see it.
4:05–4:08. Use a morning brief as the example. An automation still needs a clear brief; scheduling vague work creates repetitive vague output. Start by running the prompt once, reviewing it and only then scheduling it. Do not imply a task can silently inspect every chat or local file.

35. Three useful things.
Ready before the day starts.
4:08–4:14. Run the morning-brief prompt from /workshop/ once. Ask for public sources, publication and event dates, three items maximum and no invented news if little changed. Open one cited source and compare it with the summary. Employment or policy claims need direct authoritative sources; do not present a news brief as legal advice. Explain that no employee records or corporate inbox access are needed for this version. Fallback: use a prepared public article and show the same summary structure, clearly labeling it a single-source example rather than fresh scheduled research.

36. Try it once.
Then ask for it again tomorrow.
4:14–4:19. Participants who want to can ask: Run this brief each weekday at 7:30 am America/Chicago. Show me the saved schedule. Demonstrate the task notification settings. ChatGPT task-completion emails do not require connecting the company's inbox or adding Resend. Teach the email as notification that the result is ready; do not promise the full formatted report in the email. Completion check: either a visibly saved schedule with the right zone, or a saved prompt ready to rerun. Fallback: save the prompt and demonstrate the scheduled version on TJ's available setup; be clear which account actually has the task.

37. Turn an interesting link
into a useful decision.
4:19–4:25. Use a public article from the morning-brief demo. Copy the ideas-manager instructions from /workshop/ into a dedicated chat. Add one sentence about the participant's role. Ask for a short answer and one next action, with uncertainty stated. If the link cannot be read, paste an excerpt and tell the assistant its limits. Completion check: a recommendation tied to the participant's stated role and the article's evidence, not generic praise. Fallback: use a public excerpt with its source title and date.

38. Now make it yours
4:25–4:28. Protect these final thirty-five minutes. Invite questions as people build. Offer three easy choices: onboarding checklist, employee FAQ from a fictional policy, or a training outline. People may continue the assessment or scheduling exercise. Choose a small deliverable someone can inspect today.

39. One task. One useful result.
4:28–4:43. Fifteen-minute work sprint. Start with: Help me make [deliverable] for [audience]. These are the facts. Ask about missing details before drafting. At minute five check everyone has a first result. At minute ten ask for a targeted revision. Walk the room and solve one blocking question per person. Completion check: a visible result and the prompt or instruction block needed to repeat it. Fallback for anyone stuck: draft a fictional first-day welcome checklist with owner, action and missing information columns.

40. Would you put your name on it?
4:43–4:48. Pair up and inspect one output. Prompt: Compare this result with my instructions and source facts. Identify the three most important problems before revising. AI's self-check is useful but not proof; the participant still inspects the source. Completion check: one corrected issue or a concrete explanation of what was checked. Fallback: use the projector to review a volunteer's fictional result together.

41. Show us the useful part.
4:48–4:56. Invite four volunteers for about one minute each and use the remaining four minutes for questions. Keep sharing optional. Show only fictional or approved material. Ask the group to name an instruction they want to borrow. If no one volunteers, ask pairs to share verbally rather than exposing screens. Completion check: participants have heard at least two concrete examples from peers.

42. Start with one task.
Keep the useful part.
4:56–5:00. Ask everyone to name the next task they will try, privately or aloud. Remind them that /workshop/ has the reusable prompts, fictional sample cases and today's materials. Answer the final question, thank Tony and the team, and leave the companion link visible. Do not promise unattended access or finished integrations that were only demonstrated as drafts. The win is a useful result and a repeatable way to ask for it.