Prompt Is Also Interface Design

In a research-oriented AI app such as Deep Research, a prompt is not just a hidden string on the backend. It is part of the product interface.

The user’s research question, the way the system decomposes work, the structure of the generated report, and the source citation rules are all shaped by prompts. If the prompt is unstable, the frontend can be clear and polished, but the output will still feel unreliable.

Prompt as Product Boundary

I prefer to think of prompts as interface contracts:

• the input should make the user’s research intent explicit;
• the output should be constrained into a readable and reusable structure;
• unknowns should be exposed as uncertainty instead of guessed away;
• sources and conclusions should stay connected;
• failure states should be designed instead of reduced to “please try again.”

This is close to traditional UI design. Buttons, forms, empty states, and error messages all define the relationship between a user and a system. Prompts do the same thing, except the boundary is between human intent and model behavior.

Specific Problems in a Research Workflow

Deep Research’s backend has to coordinate model calls, search tools, state flow, and persisted data. Prompt design here cannot optimize for a single response. It has to fit into the full workflow:

1. The search stage should produce useful queries, not premature conclusions.
2. The analysis stage should separate facts, inferences, and missing evidence.
3. The report stage should keep structure stable enough for the frontend to render.
4. If the material is insufficient, the model should show the gap instead of fabricating a complete answer.

Prompt engineering is not about making wording clever. It is about turning the most unstable layer of an AI system into a maintainable product boundary.