Introducing the CSD AI Launchpad

Building websites and apps with AI

Building software is no longer a technical skill. It is a matter of telling the AI what you want, clearly enough that it can do the work.

Center of Excellence in Systems Engineering · Purdue University Fort Wayne

Learning objectives

By the end of this session, you will be able to

  1. Describe what an entrepreneur can build with AI today, from a website or an app prototype to the systems that run a business.
  2. Learn why AI is a brilliant consultant with amnesia, and how to tell it what you want and give it what it needs to do the work.
  3. Learn to tell AI what you want: who it is for, what it should do, what good looks like, and how we will know it works. This is managing, not coding.
  4. Experience how AI can be used to build workflows (systems) instead of just doing tasks.
  5. Tell a chatbot from a harness, such as Claude Code, which reads and writes real files, checks its own work, and keeps its memory in the folder.
  6. Build a real website or app in the Launchpad, get coached on your prompt, share a live link, and know how to keep going after class.
Everything in the next hour serves one idea: the people who get great results from AI are the ones who can say clearly what they want.
What an entrepreneur can do alone now

You can build the whole digital side of a business in a few weeks, for almost nothing

What you can build yourself

  • A real website that takes orders and sign-ups.
  • A real application, such as bookings, memberships, or a customer portal.
  • An entire system to run the business: inventory, sales, invoicing, point of sale.
  • And the thinking work around them: the marketing, the product design, the processes.

It is happening everywhere

  • One person with no coding background has shipped three live software products using Claude Code.
  • Founders now get a first version out in five to seven days, work that used to take a hired team two or three months.
  • Eight million people build software on Lovable without writing code, and Claude Code alone is a $2.5 billion a year business.
  • As Andrej Karpathy put it, the hottest new programming language is English.
The AI costs a few dollars a month and the hosting is free or close to it. The barrier used to be whether you could code. Now it is whether you can say what you want.

Sources (2026): the Claude Code Guide for Startups at claude.com; "From Zero Code to 3 Live SaaS" on dev.to; founder playbooks from lowcode.agency and stormy.ai; Forbes and TechCrunch on Lovable.

How most people use AI today

As a chatbot: a brilliant consultant with amnesia

One task at a time

You type a question. It answers. You copy the answer somewhere, fix it up yourself, and next time you start over. That is fine for an email or a summary. It is useless for running anything.

Starting from zero, every time

It knows nothing about you, your business, or what you told it yesterday, unless that is handed to it right now. Whatever gets handed over in that moment is called the context.

The model didn't change. The context did.

Today's idea: stop giving AI tasks and start building a system that does work for you. Give it the context and the tools it needs, and write them down so they are there every single time.
An experiment: the same AI, the same afternoon, two prompts
Prompt
make a logo for my business
The AI asked backWhat is the business called? Who is it for? What feel, and what colors?
Second try
a logo for a coffee shop
A polished vintage badge logo reading Aurora Coffee Roasters, est. 2021, a business the AI invented because the prompt never said whose logo it was
The AI is drawing…
Invented: “Aurora Coffee Roasters, est. 2021”
Prompt
Logo for 'Mastodon Coffee Cart', a student-run espresso cart on the PFW campus. Who it's for: busy students and staff between classes. Feel: warm, friendly, a little playful, trustworthy. Style: flat vector, two colors — espresso brown and PFW gold — white background, no gradients. Composition: round badge; a mastodon silhouette holding a coffee cup; the words 'MASTODON COFFEE CART' around the badge. Must read at 1 inch wide (a cup sticker). No slogans.
A clean round badge logo: a brown mastodon silhouette holding a coffee cup on a gold circle, with the words Mastodon Coffee Cart around the rim
The AI is drawing…
Exactly what we asked for
The first fix: say what you want

Four things every prompt needs, and it is managing, not coding

1. Who is it for?The customer. For the coffee cart, it was busy students between classes.
2. What should they be able to do?The job. Order quickly, see prices, book a first walk.
3. What does good look like?The tone, the colors, and an example you admire.
4. How will we know it works?The check. Someone can submit the form and see a thank-you.
This is exactly how you would brief a new employee. You don't need to know how they will do the work. You need to be clear about what you want, who it is for, and what done looks like. Good input produces good output.
The second fix: build a system instead of doing a task

Watch a workflow get built, one turn at a time

We asked a plain AI for "a document". The next slide replays that real conversation, turn by turn, exactly as it happened. Nothing has been edited or tidied up. After each turn, answer before I advance: what would we say next?

The setup was as plain as it gets: a fresh assistant, an empty folder, and no instructions of any kind. That is exactly what a new user gets.
The workflow, replayed · turn 7 of 7
We said
can you put together a document that explains how to change a flat tire
The AI said

"Per the workflow we set up, a few quick questions so I don't invent anything. Should it cover just a standard spare, or run-flats and cars with no spare? Torque values and tire pressures vary by car, so are you fine with generic guidance rather than made-up figures? Who is it for and what tone? Same one-page style, or should there be a safety section at the top? Anything to leave out?"

We gave it five short answers and it delivered a finished, checked PDF on the first try.
A one-page flat-tire guide in the same style, with a before-you-start safety section, produced first try
This was the same kind of request as turn 1, and the result is completely different. The difference is that the packet now carried everything we had taught it.
What you just watched has a name: a harness
Left: a chatbot loop, where you type, it answers text, and you copy it somewhere and check it yourself. Right: Claude Code, where you steer; it reads your files, writes the files, checks its own work and repeats; its memory is the folder and CLAUDE.md is read every time.
HandsIt creates and edits the actual files, so the website is real rather than a description of one.
EyesIt runs checks on its own work, such as "works on a phone", and fixes what fails.
MemoryCLAUDE.md holds your brief, your rules and what is built so far, loaded on every prompt.
ReachIt deploys to the live website or app, then opens the live address and checks that too.
Reach is what closes the loop A human in the loop versus the AI closing the loop itself Left: with you in the loop, the AI writes a change, then you copy it in and deploy to the live website or app, you open the site and check, and you type back what is wrong. Three of the four steps are yours, so you get two or three laps. Right: with reach, the AI edits the files, deploys to the live website or app, checks it against your rules, and if it is not good yet it fixes and goes again, as many laps as it takes. When it is good it reports back and you review. You in the loop every lap needs your hands Lap 2 AI writes the change You copy it in, you deploy to the live website/app You open the site, you check it You type back what is wrong ⏳ waiting for you… Three of the four steps are yours. The AI waits between each one. You get two or three laps before you run out of evening. You out of the loop reach: it deploys to the live website/app and checks it itself Lap 6 ✓ It edits the files It deploys to the live website/app ✓ checks pass It checks the live site against your rules Not good yet? It fixes and goes again Good? It reports back. You review the finished result. As many laps as it takes, in minutes, while you do other things.
Every step a person has to do by hand is a step the AI cannot repeat on its own. Out of the loop, it keeps going, build, ship, check, fix, until your definition of "good" is met. Your job becomes defining "good".
What the harness packages and sends on every single call
Every prompt sends the whole packet, assembled for you, every single time
1 · Standing instructionsCLAUDE.md · your rules · what tools it may useloaded automatically
2 · The files in the folderthe pages it wrote last time · your brief · your picturesread on demand
3 · The conversation so farwhat you asked before · what it did · what the checks saidresent each turn
4 · Your new messagewho · what · what good looks like · how we'll knowthe only part you type
one packet→
The model brilliant · fast · tireless …and it remembers nothing from one packet to the next
📦 the packet

The model didn't change. The context did.

A chatbot only has layer 4. Claude Code sends all four, and that is the difference.
The same pattern, with a different folder

This works for more than websites

  • Marketing. The folder holds your brand voice, your customers, and every post you have approved. The AI drafts the next one in your voice and checks it against your rules.
  • Business coaching. The folder holds your plan, your numbers, and last week's review. Every Monday you ask what to focus on, and it answers from your actual situation.
  • Tech support. The folder holds your product docs and past answers. It drafts replies to customers, and you approve them before they go out.
  • App design. The folder holds who your users are, what they need to do, and screens you admire. It sketches the screens and builds a clickable prototype.
  • Market research. The folder holds your questions and your sources. It summarizes with citations, and the next session picks up where you left off.
In every case the model is the same. What changes is the context you keep in the folder and the checks you ask it to run.
Live: write a request, let the Coach review it, improve it together, then build it

Checking the demo account…

0:00

    csd-launchpad.sysdesign.org
    The page appears here as the AI writes it.

    It's live. Open it on your phone.

    Now watch. This video was made by AI.

    The script, the voice, the visuals and the editing were all done by AI, with no crew and no studio. Now that you know how this works, ask yourself what someone had to tell the AI to get this, and what you would have to tell it to get yours.

    Your turn, for the next twenty minutes

    AI Launchpad

    1. Go to csd-launchpad.sysdesign.org and create an account. The class code is on the board.
    2. Pitch it. Start with "Build me a website…" or "Build me an app…", describe your idea, then ask the Coach. It tells you what's missing. Add what you can, or accept it as is and move on.
    3. Write one rule. Finish the sentence "The builder must always…"
    4. Build the first version. Ask for the home page, and read what the Coach predicts before you build.
    5. Open your live link and text it to someone.
    QR code that opens csd-launchpad.sysdesign.org
    Scan to startcsd-launchpad.sysdesign.org
    Launchpad is a friendly, structured front door to Claude Code. It is the same AI and the same four habits, with the training wheels on. What you make is real source files that you keep.

    If you don't have a business idea yet, borrow one: a campus coffee cart, a dog-walking service, tutoring, or a vintage clothing stall.

    Report out, five minutes

    Three volunteers, one minute each

    • Show us your link on the screen.
    • What did the Coach catch that you had left out?
    • What did the Builder get wrong, and what would you tell it next?

    The people who get great results from AI aren't better with computers. They are better at saying what they want.

    Where you go from here

    Take it further, if you want to

    1. Download your source code from Launchpad. Inside is a CLAUDE.md with your brief, your rules and the four habits. It is your project's memory.
    2. Sign up at claude.ai. It is free to start. Paste your CLAUDE.md in and keep refining.
    3. Try Claude Code, about $20 a month. Open a terminal in your folder, type claude, and it picks up exactly where class left off.
    4. Put it online. A website can go on GitHub Pages for free, at a github.io address. An app can run on fly.io for about $1 to $3 a month, because it sleeps when nobody is using it. Your own domain name from Porkbun is about $12 a year. Then send the link to three real customers.

    Other harnesses, as of fall 2026: Codex CLI, Gemini CLI, Cursor, GitHub Copilot, and in the browser Lovable, Replit and Bolt. Names and prices change monthly. The four habits work in every one of them.

    Stuck or curious? Ask me any time: smitjj09@pfw.edu · Joe Smith, Center of Excellence in Systems Engineering.
    PDF Back to AI Launchpad