Lectures

We examine why the system‑prompt in a ChatGPT App is not “persona and style,” but a behaviour contract: role and boundaries of responsibility, rules for invoking tools, safety and user consent, data minimisation, as well as the practice of versioning prompts in your codebase (down to diff). Using the sample training app GiftGenius, we assemble a complete specification prompt and discuss common mistakes.

We break down how to turn a systemPrompt into a UX contract for ChatGPT Apps: when the assistant should announce launching GiftGenius, when to deliberately not open a UI and answer in text, how to conduct the dialog after a widget (summary + follow-up), and how to honor requests for “no apps.” Plus practical prompt templates, an example call to sendFollowUpMessage, and recommendations on structuring instructions.

We break down a “three‑layer defense” against hallucinations in ChatGPT Apps: global rules (layer 1) via a pseudo‑ system‑prompt, a formal contract at the tool description/ inputSchema level (layer 2), and follow‑up UX patterns for empty/erroneous results (layer 3). Using the GiftGenius catalog as an example, we design a consistent “system ↔ tools ↔ follow‑ups” contract so the model does not invent items, prices, or actions outside the backend’s data.

In this module, we break down how to describe model behavior for the ChatGPT App through concrete use-cases and JTBD, and then lock in the expected tool-calling decisions via a golden prompt set. Using GiftGenius as an example, you’ll see how to turn a “feature list” into scenarios, scenarios into rules for the system-prompt and tool descriptions, and those rules into a verifiable engineering artifact in the repository (JSONL + a simple API route in Next.js).

In this material, we look at how to turn a good system-prompt into a real working app through solid tool ( tools) and metadata design: how to slice functionality by the “one task — one tool” principle, how to design inputSchema/ outputSchema, how naming and descriptions affect discovery/routing, why annotations like readOnlyHint and destructiveHint are needed, and why all this resembles “Tool Call SEO.”