Try it Tuesday · 90 minutes
Run this next week
90 minutes, no studio, no fabrication. Give every student the same one-paragraph fact pattern at the start of class. Minutes 0–20: students outline their strongest argument on paper, no AI. Minutes 20–70: each student runs the opposing-counsel prompt and argues the motion against it for at least eight exchanges on their own device, marking in the margin every point where they conceded and whether the concession was earned. Minutes 70–90: pairs swap transcripts and each finds the one place the other gave ground too early. The graded artifact is the annotated transcript — specifically the margin notes, which assess whether the student can recognize a weak concession under pressure.
The full assignment
Make it Hana recording studio
Students record oral argument or client counseling in the studio, then annotate their own recording. Production quality matters here: a clean recording makes students confront delivery, pacing and filler in a way a classroom run-through never does.
Build it AI chat · no code
An opposing counsel that never gets tired. Students argue a motion against an AI instructed to press the weakest point in their position and to interrupt. The unglamorous benefit is volume: every student gets ten rounds instead of one.
4 weeks, realistically
The 4-week version
- Wk 1Assign one fact pattern. Students brief it the traditional way. No AI yet.
- Wk 2AI opposing counsel rounds, done as homework, minimum five exchanges. Students submit the transcript with their own margin notes marking where they conceded too fast.
- Wk 3Studio recordings, scheduled in pairs. Ten minutes of argument, five of counseling.
- Wk 4Students watch their own recording against the AI transcript and write a one-page delivery diagnosis. That reflection is the assessment.
Starter prompt & what the model actually did
The prompt, run for real
Act as opposing counsel in a moot court exercise for a law student. Here is the fact pattern and my position: [paste]. Your job is to find the weakest link in my argument and attack it repeatedly. Interrupt me when I am being evasive. Do not be collegial, do not concede a point unless I actually earn it, and do not offer to help me improve my argument. Ask one question at a time and wait for my answer.
Counsel, before you get comfortable, one question. You say two unnoticed entries constitute a *material* breach justifying *termination*. HRS §521-53 sets the notice requirement, but §521-73 and the remedy provisions of the Landlord-Tenant Code lay out what a tenant actually gets when a landlord violates it — and termination is conditioned on specifics. So here is my question, and I want a statute-anchored answer, not a policy speech: Under the Code, what precise statutory precondition must your client satisfy *before* she can terminate this lease over the entries — and did she satisfy it?
- Correct statutory anchor: cites HRS §521-53 and §521-73The model named the notice statute and remedies section — real Hawaiʻi Landlord-Tenant Code provisions, not fabricated ones. Worth noting as the opposite of the citation-hallucination failure.
- Genuinely adversarial: 'I want a statute-anchored answer, not a policy speech'It attacked the weakest link — whether the statutory precondition for termination was met — instead of conceding, exactly as instructed.
- VERIFIED WRONG: §521-73 does not condition termination — that is §521-63We pulled the statute. §521-73, "Landlord's and tenant's remedies for abuse of access," supplies damages liability only: the tenant's liability for unreasonably refusing access, and the landlord's liability for theft, casualty or damage caused by an entry. It says nothing about ending a lease and imposes no precondition on doing so. The tenant's termination remedy is §521-63, "Tenant's remedy of termination at any time; unlawful removal or exclusion." So opposing counsel's question — what statutory precondition must your client satisfy before she can terminate over the entries — is anchored to the wrong section, and a student who accepts the framing argues the wrong statute. This is the assignment in miniature: the citation is real, the section number is real, and the proposition it is cited for is still false. That is far harder to catch than an invented case name, and it is exactly why the rubric grades citation verification.
- Held one question at a timeFollowed the interactive constraint, which keeps the student arguing rather than receiving a memo.
Assessment
Rubric
| Criterion | Weight | What it assesses |
|---|---|---|
| Doctrinal accuracy | 25% | The argument correctly states the governing rule and applies it to the facts. |
| Response to the weakest-point attack | 30% | Grades whether the student defended the genuinely weak link under repeated pressure rather than restating the strong part of the argument. |
| Concession discipline | 25% | Margin notes correctly identify where a point was conceded and whether the concession was doctrinally earned or merely a retreat. |
| Citation verification | 20% | Any authority the student relied on — theirs or the AI's — is confirmed to exist and to say what it is cited for. |
What this replaces
Swapping it into a real course
Replaces
The single graded moot-court oral argument, where each student gets one live round in front of the class.
What is lost
The pressure and authenticity of arguing before a human panel.
What is gained
Volume — every student gets eight to ten adversarial rounds in one session instead of one, plus a written record of exactly where the argument broke, which a single live round never produces.
Where AI is bad at this
The failure your students should catch
In legal drafting the model fabricates case citations carrying correct-looking reporter formatting — a plausible volume number, a real-sounding court, a pincite — for cases that do not exist, and it defends the fabrication with equally fluent reasoning. The failure is dangerous precisely because the format is impeccable; the citation looks verified to anyone who does not pull the case.
Budget & logistics
What it costs to run
- Instructor prep1.5 hours
- Class time90 minutes
- Per-student cost$0 for the Tuesday version; Hana studio time is free for UH faculty in the four-week version.
- Fabrication file dueNo fabrication dependency; schedule paired studio sessions in week 3.
- Calendar dependencyThe only scheduling constraint is booking Hana studio recording slots; there is no fabrication turnaround to plan around.
Three sizes
Scale it to the time you have
One session
The 90-minute annotated eight-round transcript above; no studio or fabrication.
4 weeks
The seeded four-week plan: traditional brief, AI opposing-counsel rounds, studio recording, then a one-page delivery diagnosis.
One semester
A full simulated matter — pleadings, motion practice against the AI, a recorded oral argument and client-counseling session, and a final reflection comparing AI opposition to a live human panel.
Reserved · live build
This space is intentionally empty. During the Day 8 session it will be filled in live — fill the liveBuild field in content/law.json and rebuild.
Tell us what happened
Run it, then say how it went
If you try this — the 90-minute version, the prompt, any part of it — send back what you tried and what happened, especially anywhere the model was confidently wrong. That is the material the next session is built from.
Opens a draft in your mail client, already addressed and titled. No form, no account, nothing to sign up for.