Everything in this course has been preparation for one question: can you build an AI product that actually works?
Not a demo. Not a proof of concept. A real product — with prompts that produce reliable output, safeguards that catch failures, handoff points where humans stay in control, and a specification clear enough that a developer or stakeholder can understand exactly what they're building.
Most people who talk about AI products stop at the interesting part: "What if we used AI to do X?" That's 2% of the work. The other 98% is designing the workflow, writing the prompts, testing edge cases, defining where humans intervene, and documenting it in a way that doesn't fall apart the moment you hand it to someone else.
This module is your capstone. You're building a complete product specification. Not an idea. A spec.
Three founders walk out of a two-hour meeting with one decision: they're going to build an AI product. Not a feature. A product.
They all feel the pressure. They know that the market moves fast. They know that the first version doesn't have to be perfect — just useful enough to get users and feedback. But they also know that shipping broken tools destroys trust, that poorly designed systems waste people's time, and that if they're going to put something in users' hands, it needs to work.
So they're asking you: what does this product actually do? What does the user experience look like? How do you ensure that the output is reliable — or at least that when it's wrong, the user catches it before they make a decision based on it? What does it take to ship this?
You get to choose the product. Your job is to design it completely — not just the idea, but the workflow, the prompts, the safeguards, the handoff points, all of it. You're writing a spec that the founders can hand to a developer (or to an AI) and say: "This is what we're building."
You have four hours. You're solving for real, not perfect.
Every AI product needs five things. Leave one out, and either the developers can't build it or users will break it in the first week.
What is the user trying to do? What are the constraints — time pressure, stakes, access to information? Be specific. Not "help writers." "Help a freelance journalist write a 800-word investigative article in 4 hours, starting from raw notes."
Map the entire user journey from input to decision to output. What does the user see first? Where do they provide information? When and where does AI do work? When do humans review? What does the final output look like? Make it visible as a sequence of steps.
Write the actual prompts the system will use. Don't say "summarize the data" — write the summary prompt. Include context, the role the AI plays, what constraints it has, what format it returns. Each stage of the workflow needs a specific prompt.
Where can the system fail? What catches that failure? Is there a human in the loop, a confidence threshold, an automated check, a manual review step? Specify the safeguard — what it looks for, how it triggers, what the fallback is if it fails.
How do you know the product is working? Is it user satisfaction? Output quality? Time saved? Accuracy of AI suggestions? Specify one or two metrics that matter, and how you'll measure them in the first month.
A complete spec is never perfect. But it's complete enough that someone else can build from it and know what they're building. That's the standard.
The gap between "I have an idea for an AI product" and "here's the spec" is where most products die.
The idea sounds good: "AI coach for public speaking." Then you have to answer hard questions. Which kind of feedback? Just technical (pacing, filler words)? Or emotional (confidence, empathy)? Is the coach real-time or post-recording? Can the user ask questions, or just receive feedback? Do they record video or audio or type out what they'll say?
Each choice changes the product. And none of them are wrong — but they're all tradeoffs. Your job is to make the tradeoffs explicitly and document them so that whoever builds this understands not just what to build, but why.
Answer those three questions clearly, and you're 80% of the way to a spec that works.