Talk Prep · Synthetic Demo

Inspiration

Difficult conversations are often where people want help most, but they are also where an AI system can overreach. Talk Prep explores a smaller, safer shape: help someone prepare their own opening words without pretending to know another person's motives.

What it does

Talk Prep lets a reviewer choose a fixed fictional scenario and a goal, such as expressing feelings, setting a boundary, or repairing. The app sends only a case ID and goal. The server loads synthetic facts and priorities, then uses GPT-5.6 Terra to draft grounded openings and a boundary. Every generated part must cite source IDs.

How we built it

The project is intentionally isolated from real user data. There are no uploads, accounts, personal conversation histories, or free-text prompts. The server validates the model response before rendering it. Invalid IDs, missing citations, unsafe wording, malformed output, timeouts, or model mismatches reject the card.

Challenges we ran into

The main challenge was making the demo useful while keeping it bounded. We designed the model contract, citation checks, crisis gate, kill switch, and synthetic fixtures so the AI can contribute without silently expanding beyond the evidence.

Accomplishments that we're proud of

Talk Prep is not a relationship oracle. It refuses to infer motives or predict replies. "Still unknown" is fixed server text, and boundaries can only describe actions under the user's control.

What we learned

The most interesting design work was not making the model more expressive. It was deciding where the model should stop, what the server must verify, and how the UI can make limits visible instead of hiding them.

What's next for Talk Prep

A fuller version could add more synthetic scenarios, better reviewer diagnostics, accessibility improvements, and a clearer audit view showing exactly why each output was accepted or rejected.

Built With

• openai