Clicked Gallery
What is a system prompt?
Highlighted from a real engineering doc. Explained by Clicked.
Used in a sentence
Engineering Notes · AI Systems
Researchers showed the chatbot's entire system prompt could be extracted with a single crafted question.
The reader highlighted one word in the docs. Clicked explained the technical term “system prompt” in simple terms:
Explained in three depths
●○○
Overview
A system prompt is the standing set of instructions a chatbot's operator gives the model before your conversation begins: who to be, how to answer, what to refuse. You never see it, but every reply is shaped by it. It is written in ordinary language rather than code, and the model reads it the way it reads your messages, just first, and with more authority.
●●○
Detail
A system prompt is a block of instructions a chatbot's maker places at the start of every conversation, ahead of your first message. It tells the model what it is, what tone to take, which tools it may use, what to refuse, and small facts like today's date. It is written in ordinary language, not code, because a language model follows text, and whoever writes the opening text sets the rules. Technically the system prompt simply occupies the front of the context window, the same working space your messages share, so the model reads everything as one running conversation. Deciding what fills that space is a craft of its own, called context engineering. The system prompt is its oldest tool, and retrieval-augmented generation, which pulls outside passages into the same stream, is among its newest. The placement explains both the prompt's power and its limits. The model was trained to give the opening instructions priority, but they are instructions, not physical constraints, and prompt injection is the art of coaxing a model past them. The boundary that matters most: a system prompt is not training. Training builds the model's abilities over months. The system prompt is read fresh at the start of each chat, can be rewritten in a minute, and editing it teaches the model nothing. Companies guard theirs anyway; leaked ones make the news because they show how an assistant was told to behave.
●●●
Analogy
A system prompt works like the post orders handed to a security guard. The guard's skills came from years of training: how to watch, whom to challenge, how to write a report. What the site owner controls is one laminated sheet at the desk: everyone signs in, deliveries go around the back, call the manager before opening the loading dock. Move the same guard to a new building and only the sheet changes. The guard is no better or worse at the job, just differently instructed. That is the whole relationship between a model and its system prompt. Training built the guard; the sheet runs the shift. And rewriting the sheet retrains nobody.
AI explanations may contain errors · Not professional advice
Same facts, different vibe — Slang mode 😎
●○○
Overview
Every chat you start is secretly not the start. Before your first word arrives, the model has already read a memo from the company: be helpful, be brief, avoid these topics, today's date is Tuesday. That memo is the system prompt. You are always joining the conversation mid-stream, and the other party walked in briefed. 😎
●●○
Detail
A system prompt is the memo a chatbot reads before it meets you, and the wild part is that the most important settings file in modern software is now a polite letter. Somewhere in production there is text that says please and never and always, and billions of dollars of model does its best to comply. Its best is not a guarantee. The memo is a request, not a cage, which is why strangers on the internet spend weekends talking models out of their rules, and sometimes succeed. The memos themselves have become collectibles. People extract them with crafted questions, companies patch the hole, and a few now publish theirs outright. Reading a leaked one is oddly deflating, because behind the personality sits a list of house rules, the way the wizard turned out to be a man at a control panel. And the memo is a tenant like you: it sits in the context window and pays rent in tokens like everything else. Arranging that room is a whole job now, context engineering; the memo came first, and hauling in reading material, RAG, is one of the later tricks. Swap the memo and you have a different assistant by dinner, with nothing learned and nothing remembered. The costume changed, not the person inside. And somewhere a product manager is editing the please. 😎
●●●
Analogy
A system prompt is the note parents leave the babysitter. The sitter already knows how to mind children; that skill walked in the door with her. The note on the fridge sets tonight's rules: no sugar after seven, bed by eight thirty, pizza money in the drawer, call us before letting anyone in. The kid, naturally, plays negotiator all evening. Mom said I could. That rule is only for weekdays. That is the entire security model of chatbots: the note has authority, but only as much as the sitter's willingness to keep checking it against a very persuasive eight-year-old. New house, new note, same sitter. 😎
AI explanations may contain errors · Not professional advice
Formal definition — The same term, explained the usual way
A system prompt is a predefined set of natural-language instructions supplied by an application's operator and inserted at the beginning of a language model's context ahead of user input, establishing the assistant's role, behavioral constraints, output conventions, and available capabilities. Models are typically trained to weight these instructions above conflicting user input, though adherence is probabilistic rather than guaranteed. A system prompt is distinct from training: it modifies behavior only for the session in which it is supplied and imparts no persistent change to the model.
Want Clicked to explain terms like “system prompt” directly in your browser — including on PDFs?
Add to Chrome — Free50 free Explanations · No credit card required
More from the gallery
What is Prompt Injection?
Hidden orders inside content an AI reads — and why the attacker's text wins.
What is a Context Window?
How much text an AI can hold in mind at once — and why long chats forget their start.
What is training (in AI)?
The process that builds a model by example. Chatting with it afterwards changes nothing.
What is a large language model (LLM)?
A program that predicts what text comes next, then taught separately to answer rather than ramble.
What is AGI (artificial general intelligence)?
The hypothetical AI that could handle anything a person can. It does not exist yet, and there is no agreed test for it.
What is Temperature in AI?
The dial between same-answer-every-time and creative gambles — and where products set it.