What is a system prompt?

The message you never sent that the model reads first

Every chat app sends the model more than your message. Before your words comes a block of text the app wrote: what the assistant is called, what it should and should not do, how to format answers, what today's date is, which tools it can call. That block is the system prompt. You type "summarise this"; the model receives a page of rules followed by "summarise this".

The model treats the system prompt as standing orders. A model asked to be "a terse assistant that answers in British English and never uses bullet points" will do that for the whole conversation without being reminded, because the instruction arrives again with every message. That re-sending is the detail people miss, and it's the one that costs money, which the third section prices.

You can see real ones. Anthropic publishes the system prompts it uses for Claude in its apps, with release notes when they change. Reading one is the fastest way to understand the genre: paragraphs on tone, on when to use lists, on how to handle a request it should refuse, on the knowledge cutoff, written in plain English rather than code. Others leak; the Bing chat prompt that named the assistant "Sydney" was extracted by a Stanford student in February 2023 by asking the model to repeat the text above the conversation.

Why the same question gets a different answer in a different app

Model APIs label each message with a role, and the model is trained to weigh them differently.

Two things follow from the ordering. The system prompt sets the frame, so the same user message gets a different answer inside a different frame: "write a limerick" produces a limerick under one system prompt and a refusal under a system prompt for a banking assistant. And the system prompt is the app's, so switching apps on the same model changes the answers even though the model didn't change. What is an LLM separates the model from the app; the system prompt is most of that separation.

The line is not absolute. Models are trained to follow the system prompt over the user, but a user message can still talk a model out of its instructions, which is the whole subject of what is AI jailbreaking. A system prompt is a strong default, not a lock.

What a system prompt costs, because it is sent every turn

The system prompt is input tokens, and what are tokens in AI shows input is billed on every message. A conversation of 100 turns with an 8,000-token system prompt sends that prompt 100 times: 800,000 input tokens before a single word of yours.

Price it at list. A standard message is about 1,000 tokens in and 500 out. Adding an 8,000-token instruction, which is what 32,000 characters of English comes to at four characters per token, on every one of 100 messages costs the extra input below at the OpenRouter rates in the AI model cost index, fetched 2026-08-20.

Providers cache: OpenAI discounts cached input by half, and by 90 percent on its newest models, and Anthropic bills a cache read at a tenth of the list price, so an app that sets caching up pays far less than the chart. The chart is list price, which is what the cost index tracks and what any app that skips caching pays. Long standing instructions are still the only text sent every time.

Whizi charges the other way round. A message costs a fixed number of credits set by the model, 1 on the house model, 10 on Claude Sonnet 5, 20 on Claude Opus 5, and the length of the instruction never enters the price. The same 100 messages cost 100, 1,000 or 2,000 credits with or without the instruction. How Whizi credits work has the ladder.

Which part of the system prompt you are allowed to write

Products expose a slice of the system prompt for you to fill. ChatGPT calls it custom instructions and memory; Claude has project instructions and styles; Gemini has saved info. Whatever the name, your text is appended to the product's own instructions and re-sent every turn.

In Whizi the slice is the project. A project is a folder of conversations with its own instructions, up to 32,000 characters, applied to every chat inside it and rebuilt into the prompt on every message. Pinned files ride along the same way, capped at 32,000 characters of extracted text per file and 120,000 for the whole project block. There is no account-level custom instructions field: the per-project instruction is the only user-authored instruction in the product, so a preference you want everywhere has to be pasted into every project.

The four built-in agents show the other side. The Canadian CPA, Study Mentor, Writing Copilot and Math Tutor rows each carry a fixed system prompt written into the model catalogue, which you can't edit. Picking an agent is picking a system prompt.

What belongs in one, from the published examples

The instructions that work read like a brief to a new colleague, and the published Claude prompts are the best public model of the form. Four things carry most of the weight.

1. Role and audience in one sentence. "You are a tax assistant for Canadian sole proprietors" does more than three paragraphs of adjectives.
2. Format rules stated as rules. "Answer in under 150 words. No bullet points. British spelling." Models follow flat imperatives more reliably than preferences.
3. What to do when unsure. "If the answer depends on a fact you do not have, ask one question before answering." Without this, the model guesses.
4. What to refuse, and how. Name the categories and the wording of the refusal, or the model will invent its own.

What does not belong: facts that change. A price, a model count or a date written into a system prompt is re-asserted on every turn until someone edits it, which is how a stale instruction becomes a confidently wrong answer for months. Put changing facts in a pinned file or a search, not in the standing orders, because the model's own knowledge cutoff makes them stale anyway. Prompt engineering for beginners covers the per-message side of the same craft.