A shorter system prompt for a support assistant

A support assistant often starts with a simple instruction: answer customer questions clearly. Then the exceptions arrive. Someone adds a paragraph about tone, another about refunds, and several examples of polite greetings. Soon, the system prompt is longer than most conversations it handles.

A shorter prompt is not automatically better. But a prompt that contains only necessary instructions is easier to understand, test, and maintain. The aim is to keep the assistant’s standing rules small while supplying other information where it belongs.

§ Give the system prompt one job

A system prompt supplies high-priority instructions for an assistant. It commonly defines the assistant’s role, boundaries, and response style. The user’s message supplies the immediate request. Supporting context, such as an order record or a help article, supplies facts needed for that request.

For a beginner, the useful distinction is simple: the system prompt describes how to work; context provides what to work with. Your application still needs to place each piece in the appropriate message or field for the API protocol you use.

A company’s entire returns policy usually does not need to become a permanent behavioral instruction. The assistant can receive the relevant policy as clearly labeled reference material when answering a returns question.

§ Start with a small working version

Here is a compact starting point for a support assistant:

You are a customer support assistant. Answer using the supplied help articles and account information. If those sources do not answer the question, say what is missing instead of guessing. Never claim that an account change has happened unless a tool confirms it. Keep replies brief and give a clear next step when possible.

This prompt establishes a role, an evidence boundary, an action boundary, and a style preference. It does not repeat the same idea in several forms. It also leaves changing facts, such as account details, outside the permanent instructions.

It is a starting point, not a complete production policy. Requirements around privacy, escalation, or regulated advice may need additional rules.

§ Remove repetition before removing meaning

When shortening an existing prompt, look for instructions that do the same work. “Be concise,” “avoid unnecessary detail,” and “do not write long answers” usually express one preference.

Then look for vague requests. “Always provide excellent service” sounds reasonable but gives little guidance. “If the order number is missing, ask for it before checking delivery status” describes an observable behavior.

• Keep stable rules: role, required boundaries, and essential output requirements.
• Move changing facts: prices, schedules, policies, and customer records belong in supplied context or tool results.
• Cut decorative wording: repeated introductions and broad encouragement rarely clarify the task.
• Keep useful examples: retain an example when it explains a requirement more clearly than another paragraph would.

§ Test the shorter prompt on awkward questions

Do not judge a prompt only by how neatly it reads. Save a small set of realistic questions and run them before and after each change.

Include a straightforward question, a request with missing information, a question the reference material cannot answer, and a request to perform an action. Check whether the assistant answers accurately, asks useful follow-up questions, and avoids claiming success without confirmation.

Change one instruction at a time. If removing a sentence causes a repeatable failure, restore it or replace it with a clearer rule. Prompt length is a maintenance choice, not a score to minimize.

§ Keep the experiment small

Vana API provides an OpenAI-compatible endpoint, so an existing OpenAI SDK application can use it to test prompt revisions. It bills per real token with no monthly fee. Shorter inputs can reduce token usage, although caching and model pricing affect the actual saving.

Create a Vana API account to try a small set of questions with your first prompt and a revised version. Keep the model and other settings unchanged during the comparison.

The useful result is not the shortest possible prompt. It is a small set of instructions you can explain, test, and update without losing track of what each sentence is doing.