Prep
Light on logistics, heavy on material. The three emails and the bloated report have to be written before class, and they have to feel real. Generic filler produces generic drafts and the whole triage falls flat.
Materials
- Projector, and a chatbot window big enough to read from the back.
- Three prepared workplace emails of rising difficulty, printed and shared digitally.
- One deliberately bloated ten-page document for the compression round.
- A shared doc or a wall for pinning before-and-after prompt pairs.
- Team numbers or table cards for the prompt battle.
Teacher note
Accounts to pre-stage
- The chatbot account from Module 1 is all that is needed. Check in the first ten minutes that everybody actually has one working, because a few people will have signed up and never logged back in.
- The personal-agent block needs an assistant with memory or custom instructions. Know which of the free tiers in your room supports that today, and have a fallback (a saved instructions block they paste at the top of a chat) for the ones that do not.
- mermaid.live needs no account at all.
Fifteen minutes before
- Run "write about dogs" once so you know exactly how bad tonight's version of bad is.
- Open the three emails in tabs, in order, so you are not scrolling a folder mid-block.
- Decide the prompt-battle challenge and write it on the board face down or covered.
- Check the free tier's rate limit by sending a few messages fast. It sets how many battle entries a team gets.
- Have your own two writing samples ready for the personal-agent demo, so you can model it with real text.
Wifi fallback
Triage works on paper. Print AI drafts of all three emails beforehand and have students hand-edit them, which is arguably the better version of the exercise anyway. The prompt battle becomes a write-off judged on the prompt alone, no output needed.
Depends on earlier modules
- Module 1 gave everyone an account and the predict-the-next-word demo. Tonight is where that turns into technique.
- The personal agent started tonight runs the whole course. It picks up work in every later module and becomes the custom assistant build in Module 7, and it is a ready-made capstone.
- Module 3 reuses the iterate-do-not-accept reflex on image prompts.
What they should walk out with
Big idea
The first output is a draft, never an answer. Getting good at asking is the whole skill, and the best trick it buys you is compression.
Maps to job skills: prompt engineering · AI-assisted business writing · meeting documentation · process diagramming · professional email judgment.
Teacher note
Everyone should leave with drafts they would actually use and one before-and-after pair they can point at. The humanization line lands here and it lands once: send the one-pager somebody wanted, not the generated thousand words.
Session timing
The three hours0:00 – 3:10
| Time | Share | Block |
|---|---|---|
| 0:00 – 0:10 | Warm-up: which chatbot did you pick, and why? | |
| 0:10 – 1:15 | Lecture: text generation, summarization, translation, prompt structure. | |
| 1:15 – 1:25 | Break. | |
| 1:25 – 3:00 | Lab: Say what you mean. Broken out below. | |
| 3:00 – 3:10 | Wrap-up and the homework handoff. |
The lab block, minute by minute95 min, from 1:25
| Lab clock | Share | Activity |
|---|---|---|
| 0:00 – 0:10 | Bad prompt, live: run it, mock it together, fix it one ingredient at a time. | |
| 0:10 – 0:40 | Email triage (big rock): three envelopes of rising difficulty. | |
| 0:40 – 0:55 | Start your personal agent: memory, custom instructions, two writing samples. | |
| 0:55 – 1:15 | Wander block. Discussion pods form on their own here, let them. | |
| 1:15 – 1:30 | Prompt battle: teams, one challenge, outputs projected, the room votes. | |
| 1:30 – 1:35 | Micro-sweep: three students read one line their agent now knows about their voice. |
Teacher note
The battle is last on purpose. It is loud, it is fun, and nothing thoughtful survives after it. If you are running behind, cut the wander to ten minutes rather than shortening the triage.
The lab, activity by activity
Warm-uplab 0:00 – 0:10
Bad prompt, live
Run "write about dogs" on the projector. Read the result out in the most enthusiastic voice you can manage. Let the room mock it. Then add one ingredient per round, role, context, format, audience, re-running each time, and watch it turn into something usable.
"That is what most people's first prompt looks like, and that is why most people think these things are useless. Watch what four small changes do."
First five minutes is a win: nobody has to produce anything, they just have to have an opinion about writing, which everybody does.
1. Email triage
lab 0:10 – 0:40 · together, then solo, then a pair swapActivity · big rock30 min
Three envelopes, rising interpersonal difficulty
Email one, a routine status update, drafted together with a template prompt. Email two, pushing back on a boss's request, solo, then hand-edit the tone. Email three, apologising for a missed deadline, solo, then swap with a partner and ask the only question that matters: would you actually send this?
Launch script
"Everybody has sent an email they regretted. Tonight we find the exact line where this thing stops helping and starts sounding like a stranger wearing your name."
"Swap screens. Read your partner's apology. One question: would you send that, as you, to a person who is annoyed with you?"
Template prompts for all three emails, filled in with their own specifics. Three usable drafts is a full pass.
Take the sensitive one with no template at all, then write two sentences on exactly where the AI's tone went wrong and what you changed. That critique is the transferable part.
Anticipated wrong turns
- Every draft opens with "I hope this email finds you well." Ban it out loud after the second one. It is the fastest way to show what a default sounds like.
- The apology over-apologises. Models grovel. Ask whether the tone matches what actually happened, and let them cut it in half.
- Somebody pastes a real email with real names. Catch it early and make it a teaching moment about what goes into a public tool, not a telling-off.
- The pushback email gets softened into agreement. Very common, and worth stopping the room for. If the draft no longer disagrees, it failed at its only job.
- The confident writer refuses the whole exercise ("I write my own emails"). Fine, and correct. Point them at the stretch tier: their job is to find and name where it goes wrong.
Discussion, with the answers you are steering toward
"Which of the three would you actually send unedited?"
Almost always only the status update. Drafting help scales with how little the relationship matters.
"What did you change by hand, and why?"
Usually tone, length, and one specific fact only they knew. That third one is the reason a human stays in the loop.
"Would you tell the recipient AI helped write it?"
No consensus, and do not force one. Just get people to state a rule they could defend. Module 6 turns this into a written policy.
2. Start your personal agent
lab 0:40 – 0:55 · solo, then ongoing all courseActivity15 min
Teach it who you are and what you already know
Set up memory or custom instructions: who you are, what you already know well so it stops explaining the basics, what you are working toward. Then teach it your voice by pasting two things you have written and asking it to describe your tone, correcting the description until it is right. Test it by having it draft something and grading the draft.
Launch script
"This is the one thing tonight that gets better every week you keep it. By Module 7 it has a job. For some of you it becomes the capstone."
"Two guardrails. An assistant that only ever agrees with you is broken, so tell it to keep useful disagreement. And the fact that it has adapted to you does not make it right."
Fill in the provided instructions template and save it. Four lines about yourself is a real start.
Maintain it for the rest of the course. Every module feeds it, and by Module 7 it is a custom assistant with a defined job and written guardrails.
Anticipated wrong turns
- The free tier has no memory. Give them the paste-at-the-top fallback and move on. The skill is writing the instructions, not the storage.
- The tone description is flattering nonsense. "Warm yet professional" describes nobody. Push for something falsifiable: sentence length, contractions, whether they use exclamation points.
- Somebody writes their whole life story into it. Remind them where that text lives. Two useful lines beat two pages.
- It agrees with everything immediately. That is the guardrail arriving on schedule. Have them add the disagreement instruction and re-test.
Discussion, with the answers you are steering toward
"Did its description of your writing sound like you?"
Usually about seventy percent right and thirty percent generic. The correcting is the exercise.
"What did you decide not to tell it?"
Excellent question and worth a full minute. Whatever they held back is their own information-governance line, drawn before anyone taught them the term.
3. Prompt battle
lab 1:15 – 1:30 · teams, then the room votesActivity · fun peak15 min
Same challenge, competing prompts, public vote
Teams get one creative challenge and craft a persona or system prompt for it. Outputs go up side by side on the projector, the class votes, and the winning teams explain what made their prompt work. That explanation is the part that teaches, so do not let the vote be the last word.
Launch script
"One submission per team, because the free tier will cut us off otherwise. Spend the time on the prompt, not on rerolling."
"Winners, do not tell me it was better. Tell me which words did the work."
Contribute to the team prompt and vote. Being on a team that ships one prompt is participation enough.
Write the persona and the constraints separately, predict out loud what each will do before you run it, then check your prediction against the output.
Anticipated wrong turns
- Rate limits stop a team mid-battle. The single most likely failure. One submission per team, said before they start, and keep your own account free as an emergency runner.
- One team dominates the vote every round. Vote on the prompt with the outputs covered, at least once. It changes the result more often than you would expect.
- The winning explanation is "we just asked better". Not good enough. Make them point at words.
- It runs long. It always runs long. Two rounds maximum, and the micro-sweep still happens.
Discussion, with the answers you are steering toward
"What made the winner win?"
Specificity, a named audience, and a constraint. Almost never length.
"Would the winning prompt work tomorrow, on a different model?"
Partly. Structure travels, exact phrasing does not. Worth saying out loud, because prompt folklore spreads fast.
Spotlight sweep and the wander block
Micro-sweeplab 1:30 – 1:35
Short tonight, because the battle already did the sharing. Three students each read one line their agent now knows about their voice. It takes two minutes and it makes the personal-agent thread feel real instead of like homework.
Prompts to hand the student showing
- "Read me the one line you are proudest of."
- "What did it get wrong about you at first?"
- "Show the before and after of one email."
Wander block menulab 0:55 – 1:15
Twenty minutes, both instructors circulating. Discussion pods form on their own in this module, which is a good sign. Do not break them up to enforce the menu.
- Make it shorter. Take the bloated report, define the audience and the decision the reader has to make, and compress it to one page. Hand-verify every claim that survives. This is the humanization pillar as a task.
- Mermaid diagrams at mermaid.live: ask AI to flowchart your morning routine, paste, watch it render, then break one line of syntax on purpose and fix it. A gentle first taste of the edit, run, fix loop.
- Your resume, one weak bullet at a time, watching for claims it invents on your behalf.
- Meeting notes to summary: a messy transcript in, decisions and action items out.
- Anything you actually have to write this week. Real work beats exercises, every time.
Synthesis, exit ticket, homework
Land it here
The absurd version of this technology is one person inflating three bullets into ten pages so the recipient's AI can compress it back to three bullets. Send the three bullets.
Synthesis, five minutes
Two things trade hands: one prompt that worked, and one draft somebody cut in half. Put both on the wall or in the shared doc. The compression example is the one to read out loud.
Exit ticket
Plus and delta, anonymous: one thing that worked tonight, one thing to change. Rotate the format across modules so it stays honest rather than becoming a ritual.
Homework handoff
The study guide from their own notes goes home, annotated: three places the AI genuinely helped and one place it was wrong or generic. That annotation is the assessment, not the study guide.
Preview Module 3 in one sentence: next time we make things that look real, and then we get fooled by things that are not.