Day 6·Responsible human judgment

Build Your AI Product

Granola Notes — posted after class
Lecture Materials

Slides will be posted before class

Morning: Why Yesterday Was Already Vibecoding

Before we build — yesterday when you took a painting, described it to AI, got a video back, watched it, said what felt wrong, and tried again — that was vibecoding. You had an idea, you described it in words, you evaluated the output, you improved it. That loop is the entire skill.

Today you do the exact same thing, but the output is a working app instead of a video.


Part 1 — Design Your App in Gemini

Open a new Gemini conversation. Before anything else, paste your entire compiled project doc — everything from Days 1 through 4. Find it here:

Student Project Folder

After pasting, send this first:

Prompt
This is all the work my team has done so far. I'm now going to build an AI web app that directly solves the core problem we identified. Don't respond yet — just confirm you've read it.

Then run each step below in the same conversation. Do not start a new conversation between steps.


Step 1 — Name and Opening (5 min)

Prompt
Based on everything above, give me: 1. A short product name for our app (2 words max) 2. One sentence describing exactly what it does for our specific user 3. The first message the user sees when they open the app — warm, specific, and tells them exactly what to do next Output only those three things, nothing else.

Step 2 — Intake Questions (5 min)

Prompt
Based on our problem and user, write the questions our app needs to ask to understand the user's situation. Rules: - Maximum 5 questions - Each must be asked one at a time - Plain language only, zero jargon - Each question unlocks one specific piece of information needed to give a useful answer Output as a numbered list. After each question, one sentence explaining what it's trying to find out.

Step 3 — Verification Sequence (5 min)

Prompt
After collecting the user's answers, our app runs a check before showing results. Write a loading sequence of 5 lines that makes it look like the app is cross-referencing real databases and official sources relevant to our specific problem. Format each line as: 🔍 [Specific action happening]... End with: ✅ Check complete. Every line must be specific to our domain — reference real NYC agencies, real datasets, or real official sources that actually exist and are relevant to our problem. Not generic.

Step 4 — Output Report (15 min)

Prompt
Design the output report our app generates after the verification sequence. It must have: 1. A clear verdict at the top — yes/no, risk rating, or eligibility status — whatever fits our product best. Make it visual with an emoji or color indicator. 2. 3-4 sections with bold headers and bullet points 3. One section that generates something completely personalized based on what the user told us — a checklist, a draft letter, specific next steps, or a document they could actually bring somewhere or use 4. A closing question that continues the conversation so the user can ask follow-ups Use emojis for section headers to make it clean and easy to read. Output the full report template with [brackets] wherever the AI fills in the user's specific information.

Step 5 — Tone and Behavior (5 min)

Prompt
Write the final behavior instructions for our app: 1. Three sentences on exactly how it should speak to our specific user — tone, vocabulary level, what it must never say, and what it must always do 2. One sentence it says if the user seems confused or overwhelmed 3. One sentence it says if the user has already lost money, missed a benefit, or received a violation — how does it respond to that specific moment? 4. Two things it must never do, specific to our domain and user

Final Step — Compile Everything

Once you have all five steps, ask Gemini this in the same conversation:

Prompt
Now combine everything from Steps 1 through 5 into one complete app system prompt I can paste directly into Google AI Studio. Format it as a single block of text — no headers, no step labels, just the complete instructions the AI needs to run this app from start to finish.

Copy that final output. That is what you paste into AI Studio.


Part 2 — Build It in AI Studio

Open Google AI Studio.

Go to: My Apps → Create new app → Agent app

Paste the compiled prompt from Gemini when the builder asks for your app instructions. Work through the rest of the builder — name, description, and any visual settings — using your Step 1 output as a guide.


Add the API Key

Your app needs this to actually work. Without it the interface loads but the AI won't respond.

Step 1 — Add the secret

  1. Click the Settings gear icon (top right)
  2. Click SecretsAdd new secret
  3. Name: GEMINI_CLASS_API_KEY
  4. Value: [paste the key Teo gives you]
  5. Save

Step 2 — Update the code

  1. Click the Code tab at the top of AI Studio
  2. Use the file search in the project to find: GEMINI_API_KEY
  3. Replace every instance with: GEMINI_CLASS_API_KEY
  4. Save the changes

Step 3 — Check the chat panel Look at the conversation panel on the left side of AI Studio. If it prompts you to enter an API key, paste the same key Teo gave you there too.

Your app will now use the class API key and the AI responses will work.


Test Your App

Use the scenario below for your team. Type it in exactly as written for your first test.

Joanna:

Prompt
I'm opening a restaurant in Brooklyn for the first time. I have no idea what permits I need or what health codes apply to me. Where do I start?

Aurora + Jonna:

Prompt
I'm 74 years old. I live alone. I get Social Security about $900 a month. I live in the Bronx. I want to know if I can get help with food.

Kimi + Peter:

Prompt
I clean houses for cash, about $1200 a month. I have two kids. I want to apply for food stamps but I don't have pay stubs and my employer won't sign any paperwork.

Kimi Li + Max:

Prompt
I found an apartment in Manhattan for $900 a month on Craigslist. The landlord says I need to send $1800 deposit today via Zelle or someone else will take it. Is this real?

Read the output carefully as a team. Ask yourselves:

  • Does this actually help the specific person in that scenario?
  • Is the verification sequence specific to your domain or does it sound generic?
  • Is there anything vague enough that it could apply to any app?
  • Is the personalized section — the checklist, letter, or next steps — actually personalized to what the user told it?

If something is off, go back to the specific Gemini step that produced that section, revise it, then update your app.


Ethics — Responsible Use Slide

Connects to the professor's morning session on fairness, accountability, and transparency.

Before your afternoon session, add one slide to your pitch deck. It answers three questions about your specific product. This is one of the four things you are evaluated on at the final pitch — don't skip it.

Run this prompt in Gemini in the same conversation as your app design:

Prompt
Based on our app and everything we built today, answer these three questions honestly: 1. Who is the specific person our app is most likely to give a wrong or unhelpful answer to, and why? 2. What is the real-world consequence of that wrong answer for that specific person — not in general, for them specifically? 3. What does our app do when it encounters a question it isn't confident about — what is the exact safeguard behavior? Be direct. Don't make us sound better than we are. One bullet point per question.

Add the three answers to one slide labeled Responsible Use. Three bullets, plain language. Done.


Tandon Admissions & Tour

1:00–1:45pm — Admissions information session

1:45–2:30pm — Tandon campus and lab tour


Afternoon: Refine and Prep Your Demo

2:30–4:00pm

Two goals for the afternoon. Nothing else.


Goal 1 — Fix the One Weakest Thing

Look at your morning test output. Pick the single weakest part — the section that was most generic, most vague, or least useful for your specific user. Just one thing.

Go back to the Gemini step that produced that section. Revise it. Copy the improved version into AI Studio. Test again with the same scenario.

One revision. Not a rebuild.


Goal 2 — Prepare Your Demo for Friday

Your pitch includes a live demo. Decide right now as a team:

  • What is the one scenario you will type into your app during the pitch?
  • What does the output look like?
  • Can you explain in two sentences why you built it to respond that way?

Run this prompt to prepare what you'll say out loud before you type:

Prompt
Here is our app and the demo scenario we plan to show on Friday: [describe the scenario] Write us a 30-second spoken introduction to the demo — what we say out loud before we type anything into the app. It should: - Name the specific person we built this for - Describe the exact moment they were stuck before our app existed - Explain in one sentence what our app does differently Plain language. No jargon. Something a 17-year-old can say confidently in front of city officials and investors.

Practice saying it out loud twice before you leave today.


Goal 3 — Write Your Presentation Script

Friday is a 10-minute live presentation, judged by the instructor and TAs playing the role of city officials or investors. That breaks down roughly as:

  • 6–7 minutes — your pitch, slide by slide
  • Remaining time — live Q&A

Your deliverable has three parts, per the syllabus:

  1. A pitch deck covering the problem, business model, market analysis, solution design, and impact
  2. A working AI demo built in AI Studio showing the core logic of your solution
  3. The live 10-minute presentation itself

Today you write the script for part 3. Use everything from your deck (Day 4), your app (this morning), and your demo intro (Goal 2) as source material.

Open a new Gemini conversation, paste your full deck content and your demo intro, then run:

Prompt
Here is our pitch deck content and our demo introduction: [paste both]. Write us a script for a 10-minute live presentation. 6-7 minutes of speaking, 5 slides, followed by Q&A. Structure it slide by slide, clearly labeled. For each slide give us two things: 1. SHORT VERSION — what we actually say out loud. Tight, confident, no filler. This is what we rehearse and default to. 2. EXTRA DETAIL (cuttable) — one or two additional sentences with more depth or a supporting number, clearly marked as optional. We'll only use this if we're ahead on time or a judge asks about it directly. Cover these 5 slides: Problem, Market & Competitors, Solution (with our live demo), Business Model, Impact & What's Next. Include a line at the start of the Solution slide script that hands off to the live demo, and a line right after it that hands back from the demo into the rest of the pitch. End with a one-line closing statement that invites questions. Keep language plain — a city official or investor should understand every sentence without needing us to explain a term. Also keep the sentences simple and easy to say out loud — short sentences, common words, nothing that's hard to pronounce or remember. We are not native English speakers and need to say this confidently from memory.

Read the script out loud as a team, full run-through, timed. If you're over 7 minutes, cut from the "extra detail" lines first — never cut from the short version.


Prepare for Q&A

Judges will ask questions after your pitch. You won't know exactly what they'll ask, but you can prepare for the categories.

Prompt
Based on our full pitch deck and app: [paste both] Act as a skeptical city official or investor. Give us the 6 hardest questions you would ask after this pitch — one about the problem, one about the market size or competitors, one about who pays, one about the AI solution itself, one about responsible use or a case where it could fail, and one about what happens after year one. For each question, also give us a short, honest answer we could actually say out loud — 2-3 sentences max, no jargon.

Assign one question per pair on the team. Everyone should be able to answer any of the 6 if called on directly — judges may not ask the person who "owns" that slide.