Put a model inside your product
Add one genuinely useful AI feature to something you've already built this semester. The bar: it must be better because it's AI, not AI for its own sake.
Ideas
- Your guestbook summarises the vibe of all messages
- Your API mashup explains the weather like a friend, not a table
- Your personal site has a "ask me anything" that answers from your bio
- Your poll generates a neutral summary of why people voted each way
Steps
- Write the feature as a sentence first. "When the user __, the model __, so that ___." If the sentence is unconvincing, pick another feature.
- Design the prompt. The system prompt is a design spec: tone, format, refusals, length. Draft it in a doc before any code.
- Build it with AI tools (yes, using AI to add AI). Use any LLM API with a free/cheap tier.
- Try to break it. Feed it nonsense, hostile input, other languages. Screenshot the failures.
- Design the failure states. What does the user see when the model is slow, wrong, or refuses?
Things to keep in mind
- API calls cost money per token — this is the first time your design decisions have a per-use price tag. Notice how that changes your choices.
- Disclose the AI. Users should know when they're reading a model's words.
- Keys live on the server, never in the page. Verify.
Submission
- Live URL, your system prompt (as written artifact), and your 3 best breakage screenshots with what you changed because of them.