How do you prepare input for better AI answers?

Short answer

A good AI input is more than a long prompt. The assistant's task, the current information it will consult, the missing information it will request from the user, and when it will choose a different step must be clear. If the same question gets different answers, check these four parts instead of simply extending the prompt.

Example: a customer asks for the delivery date

The instruction “Reply politely to the customer” seems sufficient. But the assistant does not know which order it is handling, where the delivery date was updated or what it may say about a delay. A well-built application takes the order number, checks the latest status in the permitted record and shares verified information. If records conflict, it routes the request to the appropriate team. Every answer does not need human approval; response permissions and exceptions are defined for the task.

The four parts of an input

1. Task: Write the task the user expects, such as “Answer the delivery question.”
2. Kaynak: Specify whether the catalog, order record or procedure is authoritative. Identify the file's owner and update date.
3. Missing information: Define what to ask if the order number is missing and what must not be shared before identity is verified.
4. Boundaries and output: Define the response's language and content. Specify which team receives an exception and what summary accompanies the handoff.

Test with real questions

Choose anonymized examples of normal requests, missing information, outdated documents, conflicting records and unauthorized actions. Write the expected behavior for each. Check that the right source is used and that the system stops when necessary, as well as the quality of the wording. Run the examples again when the source changes. OpenAI evaluation guide also recommends testing with examples.

Before you start

• Is the first type of question the assistant will handle clear?
• Who owns the authoritative information source and its updates?
• What happens when information is missing or an exception occurs?
• Which real examples will we use to measure whether the response is useful?

WhiteGate organizes information for the chosen task, connects the AI application to those sources and tests it with your team using real questions. See how we work or discuss your first application.