Game Day 1: Wrong Grades

Session: Thu Oct 22 · Game Day 1: Wrong Grades · 65 min

One spine: a user reports a symptom, and the room responds to it. Everyone has their own workspace with the incident already in it, works in pairs, and ends the session with a postmortem drafted by Claude Code and marked up by hand. Nothing is due afterwards.

Your job today is the clock and the questions. Don't explain the cause; go around the pairs and ask "what's your hypothesis, and what query would kill it?" The instructor run-sheet (private) has the answer key per workspace, the hints in order, and the repair.

Before class: every workspace staged the night before (the check is in the run-sheet). Print 13 worksheets. Have your own workspace on the projector for the share-out.

Transition: Reminders first.

Reminders

• FEEDBACK: Checkpoint 1, closed Tue Oct 13. Feedback is coming back shortly
• DUE: First Implementation Ticket, review-ready Thu Oct 29
• TODAY: Game Days and Postmortems. Everything happens in this room, and nothing is due afterwards

Sixty seconds. Say the last line twice: Nothing is due after today, because the first ticket is due next week.

Transition: What you'll be able to do by the end.

CS 4535: Software Design & Delivery

Game Day 1: Wrong Grades

©2026 Jonathan Bell, CC-BY-SA

Monday's talk covered the vocabulary. Today you use it on an incident in progress, with the evidence in a database and nobody to ask but each other.

Transition: Here's what you'll be able to do after today.

Learning Objectives

After this session, you'll be able to:

1. Find a failure that produces no error anywhere, starting from a user's report
2. Measure how far it spread, and say which question each number answers
3. Stop it from getting worse, repair the damage, and verify the repair
4. Judge a postmortem for accuracy against what you actually saw

Thirty seconds. Objective 2 is the one to say out loud: "how many people were affected" has more than one honest answer, and they'll be asked which one they measured.

Transition: How today works.

How Today Works

• Your own workspace: open <your-username>/gameday-1 in Coder. It runs a complete local Pawtograder with a seeded class, and something in it is wrong. Treat it as production
• Your tools: psql and the code. There are no dashboards and no git history
• Start here: ~/gameday-1/GAMEDAY.md has the user's report. ~/gameday-1/POSTGRES.md is a five-minute primer if you need one
• Pairs: each of you on your own workspace, comparing as you go
• Half the room starts with Claude Code, half starts without it. At minute 30, everyone can use it

Read out the pairs and which half each pair is in (the run-sheet has the split, with the two AI research students on opposite sides). If someone's workspace won't open, they pair on their partner's screen. Code-server is shared with any signed-in user.

Transition: Four rules.

Four Rules

1. Read-only until you can explain it. Use psql-ro until you can tell your partner the cause and how many were affected. Then psql-rw
2. Log everything in ~/gameday-1/timeline.md: the time, the query, and what it told you
3. Repairs run inside BEGIN. Check the row count before you COMMIT
4. Don't reset or stop the stack. Production doesn't have a reset button

Rule 2 is the one they'll skip, and the postmortem is built from it. Say that at minute 15 too.

Rule 4: a reset would make the symptom go away and destroy the evidence at the same time. Don't explain more than that.

Transition: The clock.

The Clock

Minutes	What you do
5–30	Read the report. Detect, scope, find the cause
30–40	Mitigate and repair. Block 1 before you commit
40–45	Block 2: your account
45–52	Block 3, then gameday-postmortem
52–60	Block 4: check the draft against your prediction
60–65	Two pairs share what their AI got wrong

Leave this slide up while they work. Call minute 20 out loud: "How many are affected? Shout your number, and say which question it answers." The numbers will differ, and that disagreement is the discussion.

At minute 30, tell the agents-later half they can now use Claude Code. Remind the room that block 1 is filled in before the repair commits. Walk around and check the predictions are on paper first.

Transition: At minute 40, block 2.

The Worksheet: Predict, Then Run

One page, on paper, no AI. Each prediction goes on paper before you run the thing it predicts.

1. The repair (inside BEGIN): predicted rows, what psql said, and why they differ if they do. Decide before you COMMIT
2. Your account (minute 40): the cause, how many and which question that answers, the action you took
3. Predict the AI (before gameday-postmortem): its number, and what it'll get wrong
4. Run it, then check: were you right, what it actually got wrong, and the action item you'd ship first

Show this at minute 0 with the handout of the sheet, and again at minute 40. The predictions are the point: A row count that surprises them before COMMIT is the most useful thing that can happen to them today, and a prediction about the AI keeps them from anchoring on whatever the draft says.

Transition: At minute 45, the postmortem.

Minute 45: Predict, Then Run the Postmortem

Fill in block 3 first. Then:

gameday-postmortem

• Runs Claude Code with the same prompt for everyone: read it in ~/gameday-1/POSTMORTEM-PROMPT.txt
• It reads your timeline.md, checks facts with read-only queries, and writes ~/gameday-1/POSTMORTEM.md
• It takes two to three minutes. Then block 4: check it against your prediction

Check block 3 is filled in before anyone presses enter. If someone's Claude Code isn't logged in, they run claude once to log in, or read their partner's draft. A thin timeline.md produces a thin draft. That's worth pointing out to whoever skipped rule 2. Collect sheets at minute 60.

Transition: Two pairs share.

Minute 60: What Did Your AI Get Wrong?

• Two pairs, one from each half of the room
• Leave timeline.md and POSTMORTEM.md in ~/gameday-1. We copy them after class

Pick one agents-first pair and one agents-later pair. Two minutes each: what their draft got wrong, and how they knew. Collect worksheets. Run the collect script after class.

Transition: What's next.

Up Next

Mon Oct 26: Written Standup + Clinic

Wed Oct 28: Reading Presentations 2. Two AI readings. We'll bring what today's two halves of the room found

Before then:

• First Implementation Ticket: review-ready Thu Oct 29

Close on Wednesday: The agents-first vs. agents-later comparison from today's worksheets is real data for the R11 and R12 discussion. Then the ticket: A PR open before the weekend is in good shape.