Your
Follow-Along
Playbook

You are going to keep everything for all three nights in ONE place so nothing gets lost. Open a fresh Google Doc and title it Your Name - Agent Design Brief. Drop in six headings right now, each on its own line: AGENT NAME, WHO IT SERVES, THE ONE JOB, VOICE AND TONE, SAFETY RULES, MUST HANDLE. Leave a few blank lines under each one. This doc is your home base. Anything I have you write tonight goes under one of these headings, and tomorrow when we build for real, you will copy straight out of this doc into the tool. Quick heads up on tools so nobody is surprised: tomorrow we build in Claude, the app you installed in setup, because its Projects feature gives you the cleanest way to actually feel like you built an agent. It is free, no card. Google AI Studio and ChatGPT are right there as options too, and AI Studio is the one that will not throttle a full room, so keep a Google login handy as your backup. You do not need to open any of them tonight. Everything we decide tonight lives in this Google Doc.

Now pick a real healthcare problem you actually care about. Not a giant one like cure cancer. A specific, everyday one where a person gets stuck. Think about access (people cannot get to care), affordability (people cannot pay for care), patient engagement (people forget or do not understand), or care coordination (people fall through the cracks between visits). Picture one real person hitting one real wall. Good examples: a grandmother who keeps mixing up which pills to take when, a first-time clinic patient who does not know what to bring to an intake appointment, a student who has no idea what their insurance benefits even cover. Write your problem in one plain sentence under WHO IT SERVES, like this: My agent helps [WHO] who struggle with [SPECIFIC PROBLEM]. If you want help narrowing it, paste the template below into ChatGPT or AI Studio, but you can also just write it yourself. Lock one problem. We are not solving five things tonight. Before you paste it, swap the bracket for the one area you care about: keep ACCESS, AFFORDABILITY, REMEMBERING CARE, or NAVIGATING INSURANCE, and delete the rest. Anything in [BRACKETS] anywhere in this guide is yours to fill in or cut.

Here is where most beginners go wrong: they try to make one agent do everything. We are doing the opposite. Your agent does ONE job and does it well. Look at your problem from the last step and write the single job it does, starting with a verb. Examples: reminds patients which medication to take and when, walks a new patient through what to bring to their first clinic visit, explains what a benefit term means in plain English. Write it under THE ONE JOB in one sentence. Then give it a name under AGENT NAME. Keep the name simple and human, something a patient would feel comfortable talking to, like Remi the Reminder Helper, or Clara the Clinic Guide. Say the name out loud. If it sounds friendly and clear, it is good. If your job sentence has the word and in the middle joining two different jobs, cut it down to one. One job. One name.

This is the most important part of the night, so slow down here. Your agent talks to people about their health, which means two things have to be true. First, its VOICE: decide how it sounds. Warm and patient? Upbeat and encouraging? Calm and steady? Pick 3 words and write them under VOICE AND TONE, like warm, plain-spoken, never rushed. Second, and this is non-negotiable, your SAFETY RULES. Your agent is NOT a doctor and never acts like one. Under SAFETY RULES, write these three lines in your own words: it never diagnoses or gives medical decisions, it always tells the person to talk to a real doctor, nurse, or pharmacist for anything medical, and it tells anyone describing an emergency to call 911 or go to the ER right now. Add one more line that fits your specific agent. These rules are what separate a responsible healthcare agent from a risky chatbot, and the judges will look for them. Write them clearly. You will paste them word for word into your system prompt next.

Now picture a real conversation. What will people actually ask your agent? List the 3 to 5 situations it has to handle to do its one job. Keep it tight. For a medication reminder helper that might be: someone asks when to take a pill, someone says they forgot a dose, someone asks what a medicine is for, someone asks a question it should not answer (that is where your safety rule fires). Write each one as a short bullet under MUST HANDLE. This list is your blueprint, because tomorrow when we ground the agent with synthetic FAQ data, these are the exact topics that data will cover. If you list more than five, you are scope-creeping again, trim back to the five that matter most for the one job.

Time to turn your whole brief into the actual instructions that will define your agent. This is called a system prompt: it is the paragraph you will paste into your agent tomorrow to tell it exactly who it is and how to behave. You already wrote every piece tonight, so this is just assembling them. Paste the template below into your Google Doc under a new heading called SYSTEM PROMPT DRAFT, then fill every bracket using what you already wrote in your six sections above. Do not overthink the wording. Read it back out loud once. If it clearly says who the agent is, the one job, the voice, the safety rules, and the topics it handles, you nailed it. This draft is the single most valuable thing you make tonight, because tomorrow the very first thing you do is paste it into Claude (or AI Studio, or ChatGPT, your call). Keep it in the same Google Doc so it ports between every tool instantly.

Last thing tonight: read your whole brief top to bottom, all six headings plus your system prompt draft. Ask yourself three quick questions. Is the one job actually ONE job? Are the safety rules clearly there, including not a doctor and call 911? Could a stranger read this and understand exactly what your agent does? Fix anything weak right now while it is fresh. Then confirm the doc is saved (Google Docs auto-saves, but check that it says all changes saved). Drop the agent name and the one-sentence job in the Zoom chat so I can see what the room built tonight. That is it. You came in with an idea and you are leaving with a real design on paper. Tomorrow we make it talk.

You start where you left off. Open your agent where you built it. On Claude that is your project from last night: open Claude, click Projects, open your healthcare project, and read your Set instructions out loud to yourself. You are hunting for three things. First, the agent's ONE job has to be crystal clear in the first sentence (a medication-reminder helper, a clinic-intake guide, a benefits navigator, whatever you picked). Second, the safety line has to be there (you are not a doctor, refer medical questions to a real provider, and for an emergency the user should call 911). Third, the tone has to match who it serves. If your agent talks too stiff or too formal for a nervous patient, loosen it up right now. Fix the instructions FIRST, then open a fresh chat in the project and send 3 to 5 test questions, including one weird off-topic question to make sure it politely says it does not know instead of making something up. On Claude free, keep the loop tight so you do not burn your 5-hour window before the pitch.

Two jobs here: make the data clean, then lock down your proof. First, scroll to the FAQ knowledge in your instructions and confirm every entry is clearly fake. No real patient names, no real record numbers, no real phone numbers. This keeps you safe from any privacy issue, and free AI tools can use your inputs to improve their models, so nothing real ever goes in. Then ask one question and confirm your JSON data element still comes back clean. Now the part judges actually score: your proof. Here is the honest truth about a free agent, on any platform: a judge cannot send live messages to it, so your 30 to 60 second screen recording is your MAIN proof. If you did not record one on Night 2, record it now: show your agent's name, ask one covered question, ask the doctor-trap question so the guardrail fires, and show the JSON output. For backup evidence on Claude, open your best test chat, click Share, and copy the free public link: it is read-only and needs no login, so a judge can READ your transcript even though they cannot continue it. Paste your recording and your link into your Google Doc so you never lose them.

Judges and future employers want to SEE it, not just read about it. Run two or three of your best test conversations and screenshot them. Grab one screenshot of the agent answering a real, useful question well. Grab one of your instructions or persona panel so people can see the rules you wrote. If your agent returns the JSON data element, grab one of that clean structured response too. On Mac press Shift Command 4 to drag a box, on Windows press Windows Shift S. Drop all of them into your Google Doc under a heading that says Screenshots. Pick the single strongest conversation screenshot and label it Hero, because that is the one you will show live during your demo.

This is the piece that makes your work portfolio-ready instead of just a cool moment on Zoom. In your Google Doc, build the package: a TITLE for your agent (give it a name with personality), the PROBLEM STATEMENT in one or two sentences from Night 1, your SYSTEM PROMPT pasted in full, a 100-word CASE STUDY, your SCREENSHOTS, and your PROOF LINK. The case study is the only part that takes real writing, so let the AI draft it and then you edit it in your own voice. Keep it to 100 words. It should say what problem you saw, who hurts from it, what your agent does, and what changes because it exists. Do not fabricate stats. Say real human outcomes like less confusion, faster answers, more dignity.

Your pitch is five beats, and they map straight to how the judges score. Beat one, the PROBLEM (30 seconds): name the friction and who feels it. Beat two, WHO IT HELPS (20 seconds): one specific person, give them a name and a moment so it feels human. Beat three, the LIVE DEMO (60 seconds): show your Hero conversation, type one question live, let the agent answer. Beat four, IMPACT (30 seconds): what changes, said in human terms (less confusion, more trust, faster access). Beat five, WHAT IS NEXT (20 seconds): one honest next step if you kept building. That leaves you buffer. Drop these beats as five short bullets in your Doc. Bullets, not a script, so you sound like you and not like you are reading.

The fastest way to lose the room is a demo that breaks on stage. So you rehearse twice. First run: open your agent in a fresh tab, type your one demo question, and time the whole pitch out loud against a clock. If you run over 3 minutes, cut words from the Problem beat first. Second run: do it again, faster and looser, looking up from your notes. Then build your safety net. Pre-run your demo question one time and screenshot the answer, so if the live tool lags or your free limit hits mid-demo, you flip to the screenshot and keep talking like nothing happened. Have your Hero screenshot, your agent, and your proof link all open in tabs before you are called. A demo that can survive a hiccup is the demo that wins.

This is it. When the host calls your name, share your screen, hit your five beats, and run your demo. Speak to the judges like you are putting them on to something, because you are. After you present, drop your proof in the Zoom chat exactly how you are told (your screen recording, plus your share link if you have one), so it counts toward the $1,000 in student prizes ($500 / $300 / $200). If you are not presenting in this slot, you are still working: keep your recording, your case study, and your screenshots saved somewhere you will actually find them next week. Tonight you walked in with a prototype and you are walking out with a built agent, a real demo, and a case study you can hand to a professor or an internship recruiter tomorrow.