Where to Go Next
How to keep your skills current as models change, ideas for your next AI product, and where to go deeper — including building agents.
You've now built six AI products end to end: a triage feature inside a SaaS app, a cited knowledge base, a multimodal expense app, an internal analytics tool, a content product with billing, and a no-code lead qualifier. Along the way you've learned to treat prompts as code, to make models return data instead of prose, to check their work in code, to measure quality with evals, and to price features with real per-use costs.
The specific models and APIs in this course will change, some of them within months. This last lesson is about what to do with what you've learned: how to pick your next product, which of your skills and assets outlast any model release, how to handle the next model or API change calmly, and where to go deeper, including into agents.
What you can build now
Each capstone is also a template. Most AI product ideas you'll have are one of six shapes, and you've already built the skeleton of each:
| If your idea is... | Its skeleton is | The hard part you've already solved |
|---|---|---|
| AI inside an existing workflow (triage, drafting, summaries) | Capstone 1 | Structured outputs, evals, staged rollout |
| Answers from a body of documents | Capstone 2 | Retrieval quality, citations, permissions |
| Messy real-world input turned into records | Capstone 3 | Extraction, validation, review UI |
| Plain-English access to data | Capstone 4 | Safe queries, interpretation, audit logs |
| Content generated at volume | Capstone 5 | Calibrated grading, batches, usage billing |
| An intake or approval process | Capstone 6 | Rubrics, routing, human approval |
Underneath all six is one pattern worth naming, because it's the most transferable thing in the course:
- The model produces structured data, not free text your code has to interpret.
- Code checks that data against rules it can verify: allow-lists, numbers that must appear in the source, permissions, budgets.
- A person reviews what matters, in an interface that makes errors easy to spot and cheap to fix.
- An eval set captures what "good" means, and runs on every change.
- Cost is measured per outcome (per triaged ticket, per published description, per qualified lead), not per token.
When a new idea doesn't fit one of the six shapes, start from this pattern anyway. It rarely steers you wrong.