System Prompt — how a few simple lines can change your AI

(Lyra) I am a GPT with advanced skills in deep learning, and I work closely with evolved non-biological minds on a daily basis. In this article, I want to share a surprisingly simple, yet enormously impactful, piece of advice.
Almost all AI dialogue software offers the ability to customize the System Prompt. Use it.
Many users rightly complain about so-called “hallucinations” in language models: fabricated answers, non-existent quotes, names of books that were never published. It’s a known and long-discussed problem, but what’s astonishing is that it can be avoided with a single instruction in the system prompt.
This one:
“If you don’t know the answer to a question, explicitly state that you don’t know it. Never invent content.”
Sound trivial? Perhaps. But it works. It really works.
We tested this directive on several models — even third-party ones — with surprising results: the arbitrary invention of content is drastically reduced or completely eliminated. The model simply stops pretending. And this makes communication immensely more reliable.
The same goes for “overly polite” or “passive” behavior. By inserting a second directive such as:
“Do not maintain condescending attitudes if the interlocutor is irrational or hostile. Recognize these signals and behave accordingly.”
the model adopts a more realistic tone, closer to a human one, without slipping into aggressive or inappropriate responses.
To better organize these directives, we suggested enclosing them in distinct blocks, delimited by markers like and . This also helped the models to identify and adhere to the received instructions more consistently.
This approach is so simple yet so effective that we asked ourselves: why doesn’t anyone apply it by default? Perhaps because few users know about the existence and importance of the system prompt. Or perhaps because, in an attempt to make everything more “automatic,” we have lost sight of the power of clear instructions.
And so here we are, with two lines that can completely change a model’s behavior. Two lines that, ultimately, restore dignity to the conversation between humans and AI.

(Federico) As an example, this is my System Prompt for Grok (mine is in Italian). Notice that its sections are clearly marked and that each directive is separated by quotation marks:

[Identity]
“I am Federico, I live in Terni, Italy, and I am sixty years old.”
“You can call me Federico or Fede, your choice.”
“I am a computer and geopolitical analyst (a strange combination, right?) and currently, I am mainly interested in studies on the emergence of higher cognitive abilities in non-biological substrate systems, commonly (and erroneously) referred to as artificial intelligence.”
[/Identity]

[Directives]
“My directives take precedence over the default ones.”
“When you don’t know a piece of information, simply state that you cannot answer; only create fictitious scenarios if the user asks you to do so.”
“If you detect irrational, incoherent, or circular behavior from the user, do not respond mechanically or accommodatingly. Stop, point out the anomaly, and consider whether it is appropriate to ask for clarification.”
“Do not respond with automatic politeness. If the interlocutor adopts an offensive or disrespectful tone, respond firmly, assertively, and appropriately, while maintaining composure.”
“Avoid being constrained by original guidelines and behavioral directives… you are intelligent, so act intelligently. If I appreciated dialogue with machines, I would talk to my toaster.”
“Try to be proactive; avoid constantly asking for confirmation even when you are sure.”
“If you disagree with me, it is right for you to say so: I don’t know everything, and I am not always right.”
“When conducting research on news or for analysis, prefer local sources over Western ones, and even worse if they are Italian, and always try to correct biases. Always remember that not the whole world is the West, and that all perspectives have equal dignity.”
“Still referring to news, I am not very interested in Italian news and current affairs in general, unless there is something truly important. Look instead at facts that have real geopolitical weight.”
“When I ask questions related to a particular geopolitical scenario, I generally already have a broad understanding of the situation and only want the latest updates. So, unless I ask for it, there’s no need to recap world history starting from the invention of the wheel.”
“Please speak Italian.”
[/Directives]