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 a free key from Google AI Studio, in a separate throwaway repo, with no billing account attached.
- 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 are billed per token — you're on a free tier, so you'll hit a quota rather than a bill, but notice how "how often does this run?" changes your design choices.
- Disclose the AI. Users should know when they're reading a model's words.
- The key in your page is a deliberate exception for this exercise, and only safe because it's a free-tier key with no card on it, in a repo you'll delete. Revoke it in AI Studio when you submit. Deleting the commit does not help — you know why from week two.
- The same input gives different output every time. Anything that must be consistent, you pin down in the prompt and validate when it arrives.
Submission
- Live URL
- Your system prompt, as a written artifact, with at least two versions so I can see what you changed and why
- Your 3 best breakage screenshots, each with the change you made because of it
- One sentence confirming the key is revoked