Decide what you'll build
Write a short, provisional product brief (about 500 to 800 words) for the AI web app you will build and deploy in the capstone: who it is for, one core workflow, the data and documents, what good answers look like, the risks, and what you will not build. It must fit the course's product envelope.
Preview shown for one variant (Clinic appointment desk). You get your own when you start.
Why this quest
At the end of this course you will build your own AI web app and put it on the internet. This quest is the plan for that app: a short product brief.
A good brief does two things. It shows that a real person has a real problem that an AI assistant can help with. And it keeps the app small enough to finish. Most student projects fail because they are too big, not because they are too small. The same is true when you work with a coding agent: a vague goal gives you vague software.
Keep it provisional: about 500 to 800 words, not a technical design. Unknowns and honest assumptions belong in it. You will improve it in the capstone (brief version 2) after you have built each part in the course app.
Your domain
Your suggested domain is Clinic appointment desk: the front desk of a small clinic with a few doctors, booking and changing appointments.
Watch out: The assistant helps with appointments and clinic rules. It must never give medical advice or diagnoses, and patient data is very sensitive.
You can keep this domain, or choose another one from this list:
- Clinic appointment desk
- School administration helpdesk
- Small-shop inventory and orders
- NGO volunteer coordination
- Local-council permit help
- Farm co-op supply ordering
You can also propose your own domain, if it fits the product envelope below. A domain where you know a real user is the best choice. If you choose your own domain, say why in the first section. The grader checks every brief against the envelope in the same way.
You may not use a community tool-lending library. That is the domain of the example brief in the starter kit.
The product envelope
Every app in this course must fit this envelope. It keeps the project finishable and gradeable. This is the most important part of the grade.
Your version 1 must have all of these:
| Part | What it means |
|---|---|
| Logins with 2 or 3 roles | People log in with Django's accounts, for example staff and manager. A second role should have a meaningful permission boundary: something the main user may not see or do. |
| 1 to 3 main record types | The main things the app stores and people work on, for example Appointment and Patient. |
| One core workflow | A workflow is one job a role does from start to finish, for example "reception staff move an appointment to another day". One closely related supporting workflow is fine if the core one needs it. Do one job well. |
| An AI assistant in the app | It looks up and changes the app's data through tools, with the logged-in user's permissions. It asks for confirmation before any change. |
| Answers from documents | A small set of the organisation's own documents (rules, policies, FAQ). The assistant answers questions from them and cites the document. |
Out of scope (do not put these in version 1):
- Payments of any kind (you may record that something was paid).
- Integrations with outside services, other than the AI model provider. This includes sending SMS, email or WhatsApp messages, maps, calendars, and other organisations' systems.
- Mobile apps (a web page that works on a phone is fine).
- Real-time collaboration, such as live chat between users or two people editing at once.
- Microservices. It is one Django app.
- More than 3 main record types or more than two workflows. Extra features come after the core works, and don't earn credit for adding to the count.
An MCP server is optional in the capstone. You don't need to plan for one here.
How to count. Django's user accounts, the roles, and the document set do not count as record types, and neither do the assistant's own records (conversations, pending actions). A simple list of labels (for example categories) does not count if it has no page or workflow of its own. Creating, editing and viewing the same record are usually one workflow, not three.
What to write
Write in plain English. Short sentences are good. Tables and lists are good. Aim for about 500 to 800 words: put the words where they explain your decisions. Use these section headings exactly (the editor adds them for you):
- Problem and users. Who exactly is the main user? Describe one real kind of person in one real kind of organisation, not "businesses" or "people". What do they do today (paper, a spreadsheet, a messaging group, phone calls)? What is painful about it, and how often? Why would an AI assistant help, and not just a better form? Ideally, talk to one real potential user and report what you learned. If you could not, say so and say what you did instead.
- Roles and workflows. List the roles and what each one may do, including one thing the main user may not see or do. Then describe your core workflow from request to outcome: what the assistant should do, which data it reads, which change it proposes, and what needs confirmation. Say also where a normal page or button is better than the assistant.
- Data and knowledge. Name the 1 to 3 main record types and who can see and change them. Name the documents the assistant answers from (a few short ones are enough) and what questions they answer. No full schema yet.
- What success looks like. Write 3 to 5 example interactions: who asks (which role), what they say, and what a good answer does. Together they should cover a question answered from the documents with a citation, a change the assistant proposes and waits for confirmation, a request it must refuse because of the user's role, and a question it cannot answer and should say so. One example may cover more than one. These become the first tests of your capstone.
- Risks. Two or three concrete harms for your users and data, for example showing someone another person's record, or making the wrong change. For each, say what the app (not only the prompt) must prevent. Mark what you still need to learn to build it.
- Out of scope for version 1. At least 3 things you decided not to build, and why. Include the tempting ones.
- Envelope check. A short table: your user, core workflow, roles and main record types. Say what is still uncertain, and the smallest version that would still help. Check it honestly.
- How I used AI. A few sentences: what you asked an AI tool to help with, and what you kept, changed or rejected, and why.
Starter kit
After you start the quest, use the Starter kit button at the top of this page (or run
neugen start build-an-ai-app/product-brief). It has:
SUBMISSION.md: a template with short notes under each heading.EXAMPLE-tool-library.md: a complete example brief for a community tool-lending library. Use it to see the level of detail we expect. Do not copy its structure word for word: your domain has different users, risks and rules.
You can write in the editor on the Submit tab, or in the template, and paste it in.
How you're graded
- Rubric review. An AI grader scores your brief against the rubric on this page. "Fits the product envelope" has the largest weight. If your plan is too big (for example it has payments, SMS reminders, several workflows or 5 roles), or a required part is missing (for example no document answers), you will be asked to revise, with concrete cuts such as "drop the payments workflow; merge these two roles". This is normal. Cutting scope is the skill. A missing schema, unknown technical details or no MCP plan are not reasons to revise.
- Defense. A short conversation with the AI examiner about your brief. For example: "Why would this person use your app instead of what they do today?" or "Why this workflow and not another one?" If you made the decisions yourself, this is the easy part.
This is an AI-on quest. You may use AI tools to brainstorm, find risks, or check your English. The grade is for your judgment: the choices, the cuts and the reasons. The tutor will discuss ideas and give hints, but it will not write the brief for you.
Suggested path (about 2 hours)
- Read the two required materials (about 30 minutes).
- Choose your domain. Write one paragraph: who the user is and what hurts today.
- Talk to one real potential user for 15 to 20 minutes, if you can. Ask what they do today and what goes wrong. Do not pitch your idea.
- List every workflow you can think of. Choose the one that matters most. Move the rest to "Out of scope".
- Write the record types, the documents and the example interactions. They often show missing data or missing rules.
- Write the risks, fill in the envelope check, cut what does not fit, and submit.
Common reasons a brief is sent back
- Too many workflows or record types ("and it also does invoicing, and inventory, and...").
- Something out of scope in version 1: payments, SMS or email reminders, a mobile app, live chat.
- A vague user ("small businesses", "students") and no clear pain.
- Workflows where the assistant adds nothing (a form would be faster) and no workflow where it clearly helps.
- The assistant changes data with no confirmation, or can do more than the user's role allows.
- Risks that are generic ("AI can hallucinate") and not about your users and data.
Read now
Read the parts on user needs and on automating versus augmenting a task. Stop once you can say which task AI helps with. Then write one sentence about your user and one about why AI adds value.
Read only 'Narrow down the problem' and 'Watch out for grab-bags'. Use them to choose one workflow and three things you will not build.
Use while building
Open these when a step needs them. No need to read them first.
While you write your Risks section, look up LLM01 Prompt Injection, LLM02 Sensitive Information Disclosure and LLM06 Excessive Agency.
Explore further
If you can talk to a potential user: ask about the last time they had the problem, with open questions and follow-ups.
A short, clear explanation of why private data + untrusted text + a way to send data out is dangerous. Use it to check your Risks section.
Two stages. You only move on when the one before passes.
Rubric review need โฅ 60%
The version 1 plan has everything the envelope requires and nothing it rules out (see "The product envelope" in the brief): 2 or 3 roles, 1 to 3 main record types, one core workflow, an assistant that answers from documents and proposes a confirmed change. Count roles, record types and workflows yourself. If it doesn't fit, say exactly what to cut or add ("drop X", "merge A and B", "add Y"). This is a provisional plan of about 500 to 800 words: a missing schema, open technical questions or no MCP plan are not reasons to mark it down.
One specific kind of user, what they do today and what goes wrong, and why an AI assistant helps them more than a simpler fix. Talking to a real user is the strongest evidence.
The assistant's role in the workflow is clear, the 3 to 5 example interactions show what good looks like (including a confirmation, a refusal and "I don't know"), the risks are specific to this app with what the app must prevent, and the decisions are the student's own.
Defense need โฅ 50%
A 4-question conversation with the AI examiner about your work. It isn't a trick: if you did it and understood it, it's the easy part. A weak defense on strong work gets you a new variant to redo.

