High school only. Middle school is a single 85-minute session — Session 1 — and ends there.
Multi-file projects and planning, applied directly to the capstone they pitched in Session 2. The capstone must be finished by the end of this session — Session 4 is 60 minutes of demos only.
| Time | Block | Mode |
|---|---|---|
| 0:00–0:12 | Why one file stops working | Demo + talk |
| 0:12–0:32 | Plan before you build | Talk + practice |
| 0:32–0:50 | Claude Code for real projects | Demo |
| 0:50–1:00 | Break | |
| 1:00–1:15 | Save points: copy the folder | Demo + hands-on |
| 1:15–2:30 | Capstone build | Independent |
| 2:30–2:48 | Debugging clinic | Whole group |
| 2:48–3:00 | Demo prep + wrap |
Campers continue the capstone they pitched and planned in Session 2. It has to be finished by the end of today. Session 4 is 60 minutes and is demos only — say this at the start and again at the halfway mark.
Open a single-file app that’s grown — 200+ lines. Scroll it slowly.
Ask: “If I want to change how the score is calculated, where do I look?”
Now try this live:
Change the way the score is calculated so that a wrong answer subtracts a point.
Watch what happens. Frequently the AI touches something else — the display, the reset logic — because in one long file everything is tangled together.
1. You can’t find anything. Not the AI’s problem, yours. You can’t verify what you can’t locate.
2. The AI has to hold the whole file in its head to change one line. Every AI has a limited amount it can pay attention to at once — its context window. A 500-line file eats a lot of that. A 2000-line file eats most of it, and quality drops noticeably.
“It’s not that the AI gets dumber. It’s that it’s reading a novel to answer a question about one paragraph. Give it the paragraph.”
This also explains a behavior they’ve probably already seen: long sessions get worse, and
/clear fixes it.
Split by job:
quiz/
main.py ← starts the app, wires things together
questions.py ← the question data
scoring.py ← how points work
display.py ← the window and buttons
Now “change the scoring” means the AI reads scoring.py. Faster, more accurate, and
you know where to look to check it. That second part is the point.
Campers already have a PLAN.md from Session 2. Today they pressure-test it against
what they now know about file splitting.
The single highest-leverage move in this whole track:
I want to build a quiz game with multiple choice questions, a score,
and a results screen at the end.
Don't write any code yet. Tell me what files you'd create, what goes
in each one, and what order you'd build them in.
Do this live on the projector with a fresh example. You’ll get a file breakdown and a build order back in seconds.
Now the crucial part — read it and argue with it.
“You just made five architecture decisions in ninety seconds without writing a line of code. Changing a plan is free. Changing code is not.”
Once the plan is agreed:
Let's build step 1 only: questions.py with 10 multiple choice questions.
Nothing else yet.
Run it. Check it. Then step 2. This is Session 2’s one-at-a-time rule scaled up from changes to components.
PLAN.md is what you check finished work against — you have a written list of what each
file is supposed to do, so “did it do what I asked” has an actual answer.
Everyone opens their PLAN.md and asks:
Given this plan, what files should this project have? Should any of
what you suggested be split or combined?
Change at least one thing. Save it. This is the last planning you do today — the rest of the session is building.
Demo each one live — no slides.
Look at scoring.py and tell me what it does.
In display.py, make the score label bigger.
The bug is in main.py, not the other files.
Naming the file focuses the AI and saves it from reading everything. Faster and more accurate.
/clear — start a fresh conversationLong sessions accumulate junk. Fifteen turns in, the AI is still carrying around the bug you fixed an hour ago and the feature you abandoned.
/clear
Fresh start, same folder, same files. Your code is untouched — only the conversation resets.
“Rule of thumb: when you switch to a genuinely different task,
/clearfirst.”
CLAUDE.md — notes the AI reads every timeCreate a file called CLAUDE.md in the project folder. Claude Code reads it
automatically at the start of every session.
## Quiz Game
A multiple-choice quiz in Python + tkinter.
## Files
- main.py — starts the app
- questions.py — question data, list of dicts
- scoring.py — 1 point right, -1 wrong, no negative totals
- display.py — the tkinter window
## Rules
- Python only, tkinter only. No extra libraries.
- Keep every file under 100 lines.
- Answer checking is case-insensitive.
Now you never re-explain your project. Every new session already knows.
“This is the difference between an intern you brief every morning and one who read the handbook.”
Every camper writes one for their capstone before the build block. Two minutes, and the payoff is immediate.
What would it take to add a timer to each question?
Is there anything in this project that's going to cause problems later?
Free consulting. No code changes. Use it constantly.
After a change, scroll back through the output. It shows you the edits. If you skip past that every time, you’ve given up your best verification tool for free.
This is the one piece of housekeeping that pays for itself inside the same session. Move briskly — it’s fifteen minutes and they need every one of the seventy-five that follow.
cd $HOME\Documents\Projects
cd <Name>
Copy-Item -Recurse capstone capstone-working
To restore:
cd $HOME\Documents\Projects
cd <Name>
Remove-Item -Recurse capstone
Copy-Item -Recurse capstone-working capstone
Everyone: make a copy, ask the AI to break something, restore it, confirm it’s back.
Copy every time it works. Name copies meaningfully —
capstone-scoring-works, notcapstone2.
This is Session 2’s check #3 — “did it change anything I didn’t ask for?” — made concrete.
Before a risky change, copy the folder. After the change, ask:
Compare the current version of scoring.py to what it was before my
last request. What actually changed?
You have the old copy on disk, so you can also just open both files side by side and look. Two files, two windows. That’s a diff, done by hand.
“You now have the ability to answer ‘what did it actually change?’ with certainty instead of trust. Use it before anything that scares you.”
The biggest block in the track. This is where the capstone gets finished.
CLAUDE.md describing the projectPLAN.mdbuild → run → check → copy
Say it out loud to the room every fifteen minutes.
Walk the room enforcing that last one. A copied working version means nobody demos a broken app next session.
| What you see | What to say |
|---|---|
| Building all files at once | “Stop. Which one do you have running right now? Just that one.” |
| Hasn’t copied in 25 min | “Copy your folder now.” |
| Tangled and frustrated | “When did it last work? Restore the copy and take a smaller step.” |
| One giant file again | “What are the three jobs in this file? Ask the AI to split it into three.” |
| Cruising | “Ask the AI what’s going to cause you problems later. Then fix one of them.” |
| Behind schedule | Cut a must-have. Visibly. Now, not at 2:25. |
Anyone stuck on the same problem for 20 minutes: restore the last working copy and approach it differently.
README.md explaining what it does and how to run itWhole group. Real bugs from the room — volunteers, screens up.
For each one, walk the same method out loud:
In scoring.py, when the answer is right the score goes down instead of up. Here’s the terminal output: [paste]. What’s happening?
Step 5 is the one everybody skips.
/clear and re-describe the problem fresh — the conversation may have a bad theory
stuck in itSession 4 is 60 minutes and it is entirely demos. There is no build time. Whatever runs at the end of today is what gets demoed.
Everyone, right now:
cd $HOME\Documents\Projects
cd <Name>
Copy-Item -Recurse capstone capstone-demo
That’s the version being demoed. If they tinker at home and break it, this one survives.
1. What it is, in one sentence.
2. Show it working. (Practice the exact clicks. Twice.)
3. One thing that broke and how you fixed it.
4. One thing the AI got wrong that you caught.
5. What you'd add next.
Practice the clicks. The most common demo failure is a camper clicking around live, hitting a bug they’ve never hit before, and losing their nerve. Rehearsing the path twice fixes it.
Have a fallback ready: if it crashes on stage, say what it does and keep going.
Plan before you build. Split by job. Copy every time it works.
Documents\Projects folder exists on every lab machineCopy-Item -Recurse / Remove-Item -Recurse restore cycle yourself
first — you’ll be walking campers through it under time pressureCLAUDE.md and PLAN.md templates ready to project
This Vibe Coding Curriculum by Brian Bird, created in with AI assistance, is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.