What the Capstone Actually Tests

The capstone project isn't a test of whether you remember terminology — it's a test of whether you can combine prompt engineering, workflow orchestration, integrations, evaluation, and governance into one coherent, working system for a real problem. This walkthrough shows the full arc of that process on a worked example, end to end.

Step 1: Scoping a Problem Worth Solving

Every good capstone starts with a narrow, well-defined problem — not "improve customer support with AI," but "draft first-response replies to billing questions, with a confidence-based handoff to a human agent." A scoped problem has a clear input, a clear output, and a clear definition of what "working" means before any prompt is written.

Step 2: Designing the Prompt and Workflow Structure

With the problem scoped, the design phase applies what earlier modules covered: role and structure in the prompt itself, chain-of-thought where the task genuinely requires reasoning, and a multi-step workflow if a single prompt can't cleanly handle classification, retrieval, and drafting in one call. This walkthrough traces that decision process live, showing where a single prompt was enough and where it wasn't.

Step 3: Wiring In Integrations

A capstone project connects the workflow to at least one real external system — a support platform, a spreadsheet, an internal API — using the integration patterns from earlier in the course: explicit contracts, timeouts, and a clear retry policy, rather than a fragile happy-path-only connection.

Step 4: Building the Evaluation Layer

Before calling the project "done," it needs a golden dataset and rubric — even a small one — so its quality can be measured rather than assumed. This walkthrough builds a ten-example golden dataset live and shows how a simple rubric surfaces a specific weakness in the first draft of the workflow.

Step 5: Adding Governance and Human Oversight

The final layer is responsible deployment: identifying where a wrong output could cause real harm, defining a human-in-the-loop escalation trigger for that specific risk, and documenting what the system is and isn't meant to do. Even a small capstone project should be able to answer "what happens when this is wrong," not just "what happens when this is right."

Conclusion

An intelligent workflow isn't one clever prompt — it's prompt design, orchestration, integration, evaluation, and governance working together around a clearly scoped problem. The capstone project is your chance to build that whole stack once, deliberately, instead of assembling it piece by piece under production pressure for the first time.