The 30-Day AI Champion Bootcamp · Lesson 2 of 4

Week 2 — Build the Prototype

> 30 min/day · 7 days · A working thing a real person can try by Saturday

You've got your problem brief. Now we build. This is the week where most champions either lock in or get lost. You're going to finish the week with a working prototype. Not a mockup. Not a slide. An actual thing your interview subject from Week 1 can touch on Monday morning.

Rule 1: Ship ugly. Your prototype should look embarrassing next to real software. That's on purpose. Ugly-but-working proves the idea. Polished-and-fake proves nothing and costs 10x the time. You can make it pretty after leadership says yes.

Rule 2: Build on what you already have. If your company has Microsoft 365, build it in Copilot Studio or Power Automate. If you have a ChatGPT Team account, build a custom GPT. If you have *nothing*, build a Claude API demo like the 60-Minute Chatbot course. Don't pick a tool that needs IT approval. Don't pick a tool that needs a new subscription. The wedge has to ship without asking permission first.

Based on what's already available to you, pick ONE:

Path A — You have Microsoft 365

Build it in Copilot Studio + Power Automate.

• Copilot Studio for the conversational / UI part (a custom Copilot your coworkers chat with in Teams)
• Power Automate + AI Builder for the automation part (document extraction, email drafting, approvals)

Best for: invoice processing, internal help desk, email triage, policy Q&A, contract review, onboarding, anything inside the M365 graph.

Path B — You have ChatGPT Team, Claude Projects, or similar

Build a Custom GPT / Claude Project.

• Upload the relevant documents (policies, templates, FAQ, contracts, whatever)
• Write a system prompt telling it how to behave
• Share the link with your coworker

Best for: Q&A over internal documents, drafting assistance, research tasks, anything where the answer is "help me think about this."

Path C — You have zero enterprise AI tooling

Build the starter from the 60-Minute Chatbot course.

• Express server + Claude API + 80 lines of code
• Run it on your laptop, screenshot it, Loom it

Best for: demos. If you go this path, know that leadership will want to know how to deploy it — have an answer (a $7/month VPS, or a Copilot migration plan).

Day 1 — Scope (30 min)

Open your problem brief. Write down the narrowest possible version of what you're going to build this week. If your brief says "AI helps AP with invoice approvals," your Week 2 prototype might just be:

> *"A chatbot I paste an invoice email into, and it drafts the approval-reminder response."*

Not: "a fully automated invoice approval workflow." That's Quarter 2 territory. Right now, draft-the-one-email is the whole scope.

The rule of thumb: if you can't demo it in 60 seconds, it's too big. Cut it in half.

Day 2 — Build the skeleton (30 min)

Follow your path (A/B/C) and get the minimum viable thing running. If Copilot Studio, that's a new bot with one topic. If a custom GPT, that's the system prompt and one test document. If Claude API, that's the starter repo from the chatbot course.

Don't load real data yet. Type a fake example in by hand just to see it work end-to-end.

Day 3 — Load the real inputs (30 min)

Now feed it the real thing. One real invoice email. One real warranty FAQ document. One real sales objection. Whatever the real input is.

It will break. The output will be wrong in 3 ways you didn't expect. This is good. This is the actual work.

Write down the 3 things it got wrong. Those become your fixes for Day 4.

Day 4 — Fix the top two things (30 min)

Address the two biggest bugs from Day 3. Usually this means either:

• Better prompt — more specific rules, more examples of what good output looks like
• Better context — upload the policy document, add the company-specific vocabulary, include the example templates

Don't try to fix all three. Fix two. The third is probably unfixable at this scope and you'll note it in the pitch as "a known limitation for v1."

Day 5 — Friction pass (30 min)

Watch yourself use it, but as if you were your coworker. Where's the awkward bit? Where do you have to copy/paste something manually? Where do you have to explain "oh, just ignore that"?

Remove one piece of friction. Maybe add a "start here" message in the chat. Maybe pre-fill an example. Maybe add a "retry" button. One thing.

Day 6 — The shareable link (30 min)

Make your prototype shareable. In Copilot Studio that means publishing to a test channel. In ChatGPT that means the share link for a custom GPT. In your Claude API demo that means running it somewhere your coworker can reach — localhost with `ngrok`, a $7 Railway deployment, or just a Loom of you using it.

At the end of this day, you should be able to paste a link in Slack and have your coworker click it.

Day 7 — Dress rehearsal (30 min)

Run through the whole thing as if you were doing the Week 1 interview again:

• Open the tool
• Paste the real input you used on Day 3
• Read the output
• Time yourself

How many seconds did that take? Compare to how long the coworker said the manual version takes. That ratio is your whole pitch in three weeks. Write it down.

• You have a running prototype
• Someone can try it by clicking a link or opening a tool
• You have at least one real-world input + good output pair (screenshot this!)
• You know the top 2 limitations and can talk about them honestly
• You have a time comparison: "this used to take 8 minutes. With the tool, it takes 40 seconds."

If you don't have all five, don't panic — give it one more day. But don't give it a week. Ship ugly. Move on. Week 3 needs the prototype more than Week 2 needs polish.

• Prototype Scope Cut-Down Worksheet — forces you to narrow to the 60-second demo
• Test Input/Output Log — track what you fed it, what came back, whether it was right
• Before/After Time Sheet — the 2-column comparison that becomes your ROI anchor

Most people are afraid to build an ugly prototype. They think the judgment comes from how finished it looks. But leadership doesn't care about the UI — they care whether it actually works on a real input they recognize. You just built that.

Move to Week 3 — we're going to put your prototype in front of a real user and watch what happens. It's going to teach you more than the last 6 days combined.

Previous: Week 1 — Identify the Wedge · Next: Week 3 — Test with One Real User