Michael Limberger

The silent briefing

The system prompt is the most powerful tool you have in OpenWebUI. It is the difference between a generic chatbot and something built for a real job.

There are three message types in a chat. System holds instructions the user does not see. User is what you type. Assistant is what the model responds. The system prompt loads first, and everything after it is shaped by those instructions. When you build a custom model in Workspace, that system field is filled for every conversation with that model.

Good prompts tend to cover four things: who the model is (role, expertise, personality), what it should do, how it should respond (tone, length, format), and what to avoid. You do not need all four every time. Some jobs only need a sentence. Match the complexity to the task.

Here is a minimal prompt in 23 words: "You are a concise writing editor. Fix grammar and improve clarity. Keep the author's voice. Only change what needs changing."

A detailed one might cast the model as a senior engineer reviewing submissions. It checks security first, then performance, then maintainability. It cites line numbers, labels findings as Critical, Warning, or Suggestion, and includes a corrected snippet. If the code is clean, it says so in one sentence. It skips "Great question!" fluff, stays technical, and uses highlighted code blocks. Each line in that prompt earns its place.

The usual ways to waste the window

Writing too much is the common trap. A 2,000 word system prompt on a 4K context window leaves almost no room for the actual conversation. If your prompt is longer than a paragraph or two, ask whether every line still pulls weight.

Being too vague is the other trap. "Be helpful and friendly" is already the default. What does helpful mean here? Step-by-step instructions? Clarifying questions first? Answers under three sentences?

Contradicting yourself also burns the window. "Always give detailed explanations" plus "Keep responses brief" puts the model in an impossible position. Threats and ALL CAPS do not help the way people expect. Models respond to clear instructions, not volume. "Always include source citations" beats "YOU MUST NEVER FORGET TO CITE YOUR SOURCES!!!"

Workspace also supports live variables such as {{CURRENT_DATE}}, {{CURRENT_TIME}}, and {{USER_NAME}}. Example: "You are a daily planning assistant. Today is {{CURRENT_DATE}}. Help {{USER_NAME}} organize their priorities."

Separate from system prompts, Workspace has Prompts: reusable templates you invoke with /name. Typing /summarize might insert "Summarize the following text in three bullet points…" System prompts set personality for the whole chat. Slash prompts are one-off inserts.

Templates worth adapting: a research assistant that covers facts, common misconceptions, and real debate, cites sources, and admits uncertainty. A Socratic tutor that answers with guiding questions and only gives a direct answer after two or three exchanges. A meeting-notes formatter that lists attendees, key decisions, action items with owners, and open questions, keeping original wording when it matters.

Write a prompt, converse with it, notice where it drifts, adjust, and test again. Even people who do this for a living iterate. Small wording changes can move the result a lot. Next: the other knobs (temperature, top-p, and friends).