Week 6 Tutorial Plan — Facilitator Guide

Building AI That Works · 2 hours · connected classrooms (MY + AU) · LO6

Audience: students who have read the Week 6 pre-class, declared a lane on the board, and brought ten real items from their own project. There is no lecture video this week — the deck and this seminar carry the content, so budget real time for teaching, not just coaching.

Co-teaching: two leads. Lead A is the metronome: owns the clock, calls every move to the whole room, and does not enter breakouts. Lead B owns Google Meet, cross-campus parity, and lane routing. Swap roles next week.

You need: the activity deck projected, the shared board (one doc, one row per team), the lane map, class_activity.md printed as the offline fallback, and the four lane fallback datasets linked on Moodle.

The organising principle: run the clock in common, run the content in lanes. At any given minute every team in both rooms is answering the same question. Only the answer differs by lane. This is what stops you fragmenting — Lead A never has to hold four contexts, because Lead A is only ever calling one move.

At a glance

Time	Segment	Lead	Mode
0:00–0:08	Welcome · "what are you building?" round · the diary starts today	A	Whole class
0:08–0:18	The four lanes · teams declare · B tallies live on screen	A teaches, B tallies	Whole class
0:18–0:25	Move 1 — Name your baseline (hard stop)	A	Teams → board
0:25–0:45	Move 2 — Build the eval set	B sweeps by lane	Teams
0:45–1:15	Move 3 — Build the thin slice	Both coach	Teams
1:15–1:35	Move 4 — Run both, get the number	Both coach	Teams
1:35–1:45	Move 5 — The honest sentence + the cost line	A	Teams → board
1:45–1:55	Cross-lane debrief — one team per lane reads field 5	A	Whole class
1:55–2:00	Commit diary entry 1 before leaving · to-do · next week	A, B drops links	Whole class

0:00–0:08 — Welcome (Lead A)

Fast round, one line each team: "what are you building?" Not the pitch — the one feature.
Frame the week in two sentences: "Last week you learned to judge a prediction. This week you build one — whatever kind yours is — and prove it beats the dumb thing that already works."
Say the diary line out loud now, not at 1:55: "Your build diary starts today, it is graded, and you commit entry 1 before you leave this room."

0:08–0:18 — The four lanes, and declare (A teaches, B tallies)

Teach the four lanes off the deck. Keep it to eight minutes — the detail lives in the annexes.

Then: every team declares a lane in the board's lane column. B tallies live and puts the count on screen.

Put the tally on screen and say what it means. "Twelve of you are predicting a number. Nine are reading documents. Seven are answering questions from a corpus. This class is four lanes, not one." That single number legitimises the whole week better than any slide.

The one rule that prevents ten minutes of arguing: pick the lane of the one MVP feature you are building today, not of your whole product. Roughly nine teams have genuinely hybrid products. They are two lanes and it does not matter today.

The question you will be asked: how do lanes relate to Week 1's five layers? Have the answer ready in one line — layers are what the AI is; lanes are what the task is. Then move on.

0:18–0:25 — Move 1: Name your baseline (Lead A, hard stop)

Seven minutes, whole-class, teams writing into the board. This is the highest-value block in the two hours.

The instruction: "Write down the dumb thing that already works without any AI. One line."

At 0:23 stop them and read three answers aloud — one strong, one weak, one from the other campus — and fix the weak one live. Weak answers are almost always too vague ("we'd do it manually") or too clever (already an AI system). Push for something you could run this afternoon.

Land the point: if your build can't beat this, the dumb thing is your product and you've saved yourself a term. That is a result, not a failure.

0:25–0:45 — Move 2: Build the eval set (Lead B sweeps by lane)

Twenty minutes. Teams turn their ten pre-class items into ~20 items with the right answer recorded by hand.

This is the move people skip, because it is unglamorous and building is fun. It is also the only reason anyone should believe their result.

Enforce it structurally, not socially: Move 3 does not open for a team until its eval set link is in the board. B checks the board, not the teams. Do not soften this — a team that skips it will produce a number at 1:35 that means nothing, and you will have no way to tell them so.

B works a lane sweep: all Lane 3 rooms in sequence, then all Lane 4, and so on. One lane's context at a time.

Teams with no data of their own go to the fallback dataset for their lane, linked on Moodle. No shame in it; they still make all five moves.

0:45–1:15 — Move 3: Build the thin slice (both coach)

Thirty minutes. Spec → prompt → test → iterate, on the one feature only.

Coaching lines that work: - "What's the smallest version that produces one output?" - "Run it on item 1 of your eval set. Now item 2 without editing anything." - "You're building the demo, not the product. The demo has to survive twenty items."

Watch for: teams building the interface instead of the feature; teams tuning the prompt against an eval item (contamination — tell them to pull a fresh item); Lane 1 teams disappearing into the forecasting lab.

Say this aloud at 0:45 for Lane 1: today's Lane 1 build is a lag/trend model, and today's baseline is seasonal-naive. The forecasting_lab/ is post-class optional depth. A keen student will find it and lose the hour to ARIMA if you don't say this.

1:15–1:35 — Move 4: Run both, get the number (both coach)

Twenty minutes. Run the baseline and the build over the same eval set.

The most common failure here is that the team never actually runs the baseline — they assume it would have been worse. Make them run it. The comparison is the deliverable.

Second most common: the number exists but nobody can say what it means. Ask "is that good?" until they answer in business terms.

1:35–1:45 — Move 5: The honest sentence and the cost line (Lead A)

Ten minutes, back in the room. Lane Card fields 5 and 6.

"On _ items, our scored _ against the baseline's . That is / is not enough to _, because _."

Then the cost line, which is new this week and which the cohort is weakest at: what does one use cost, and what does 100/day cost? Week 1 told them inference is cents per task; Week 2 gave them a box to write costs in; neither made them multiply. Make them multiply.

Teams whose build lost to the baseline write that sentence too, and read it in the debrief. Reward it visibly — it is the most honest thing that will happen in the room.

1:45–1:55 — Cross-lane debrief (Lead A)

One team per lane reads field 5. Put the four sentences side by side on screen.

Land it: four different problems, four different metrics, one identical shape. That shape is what Assessment 3 Criterion 3 is asking for.

1:55–2:00 — Close

Commit diary entry 1 now, in the room. Do not let this become homework; it is 90 seconds and the habit is the point.
B drops into chat: the post-class link, the lane walkthrough card for each lane, and next week's pre-class.
Next week: digital agents, avatars and twins — AI that doesn't just answer, but acts.

Cross-campus protocol

Every breakout is mixed-campus where rosters allow. B owns admissions, breakouts, and the lane map.
One visible countdown on the shared screen at all times. Lead A's move call is verbal; B types it into the main chat simultaneously so a student who drops and rejoins always knows the current move.
Parity check at each move boundary: B confirms the remote campus can hear, see the board, and has the instruction visible. If one campus has been silent for more than five minutes, B pulls a name from there.
Hand-offs are verbal and explicit so the recording and the room both track the switch.

Contingencies

Meet drops. The activity sheet works on paper. Moves 1, 2 and 5 need no tooling at all — baseline, eval set, honest sentence. Only Move 3 needs the tools. Keep the room running on paper and let B re-establish the link.
Gemini document upload fails or throttles (Lane 3). Drop to 3 documents per team and announce it at the top of the hour. If it fails outright, Lane 3 teams run the manual stopwatch baseline and design the field schema on paper — they still make all five moves and still get a number (the manual seconds-per-document).
A team has no repo. They use the shared doc today and create the repo tonight; diary entry 1 is due either way. Chase repo-less teams the day before, not in the room.
A lane has only one team. Lead A takes it personally rather than leaving it uncoached. One team per campus in the same lane → merge into a cross-campus lane table.
Few read the pre-class / nobody brought items. Do not re-teach the reading. Push everyone to the fallback datasets, shorten Move 2 to 15 minutes, and protect Move 4 and Move 5 — the number and the sentence are what the week is for.
Running long. Cut the debrief to two lanes rather than cutting Move 5. Never cut the diary commit.
Strong cohort finishing early. Send them at the second lane of their hybrid product, or at the "what breaks at 1000 uses a day?" version of the cost line.

BEX3202 — AI Applications in Business · Monash University · CRICOS 00008C