Designing prompts like product surfaces

Prompts are just another interface. We borrow common UX moves to keep them predictable.

Start with intent, not syntax

We write down the user’s intent and the constraints first. Only then do we reach for tools like few-shot examples or structured outputs.

Keep structure obvious

When the model needs to return JSON, we show the schema and a worked example. We avoid ambiguous nesting and include a rejected example so the model sees what not to do.

Copy
{
  "summary": "string",
  "decision": "accept | reject",
  "reasons": ["string"]
}

Add guardrails in layers

We blend prompt-level guardrails (clear format, constraints) with runtime guardrails (schemas, function calls, and fallbacks). Treating prompts like product surfaces keeps them consistent as the app grows.