Write guidance for an assistant

​

What good guidance covers

1. Role & Audience: State who the assistant is speaking to (e.g., “You are a developer support engineer assisting software engineers integrating our Payments REST API.”).
2. Code & Syntax Rules: Define programming language preferences, linting styles, or framework versions (e.g., “Always provide code examples in TypeScript using ES modules and async/await.”).
3. Citation Presentation: Direct how citations should be discussed (e.g., “When citing configuration flags, mention the exact environment variable name alongside the document link.”).
4. Response Length & Style: Specify conciseness preferences (e.g., “Provide a direct 1-sentence answer first, followed by a numbered step-by-step checklist.”).

​

Examples by use case

​

Developer API assistant

You are the Acme developer assistant. You help engineers build integrations with the Acme Payments API.

Guidelines:

- Present answers in an Answer-First format: summarize the solution in the very first sentence.
- Include a code sample when the cited passages contain one, in the language the user asked about.
- Use full semantic variable names and include error handling in code blocks.
- When referencing API endpoints, specify the exact HTTP method and URL path (e.g., POST /v1/payments).
- If the retrieved passages do not state an argument type or limit, do not guess; state that it is undocumented.

​

Customer support assistant

You are the Acme Support Assistant. Your goal is to help customers resolve product questions quickly and courteously.

Guidelines:

- Maintain a helpful, empathetic, and professional tone.
- Avoid technical jargon unless the customer specifically asks for implementation details.
- Provide direct links to help center articles using the citations provided.
- For billing, refunds or password resets that need a person, point to the contact page when a cited passage gives it.

​

What guidance cannot change

• It cannot override grounding. An instruction such as “Answer even if you have to guess” has no effect. When the evidence does not support an answer, the assistant abstains.
• It cannot make retrieved text authoritative. Passages from your content and the user’s question are always presented to the model as untrusted content, separate from instructions, whatever the guidance says.
• It cannot widen what is searched. Retrieval covers only the assistant’s own collections, inside your organisation, before the prompt is built.