System prompt

Chat models receive messages with roles. The system message comes first and describes how the assistant should behave: that it is a patient math tutor, that it answers in French, that it keeps replies under 100 words, that it returns JSON. User and assistant messages follow. Models are trained to give the system message more weight than ordinary text, but it is still text, and a model can drift from it or misread it, especially in long conversations.

A useful system prompt is specific. It states who the assistant is, who it is talking to, what a good answer looks like and what to do when information is missing. Typical ingredients are a persona, an audience, a tone, a length limit, an output schema and two sample exchanges, and examples of the desired format work better than abstract adjectives. Rules are clearer as positive instructions, such as “ask one clarifying question first”, than as long lists of prohibitions. Keep it short enough that nothing contradicts anything else, and test it with awkward inputs.

A system prompt is not a security boundary. Users can sometimes persuade a model to reveal it or to act against it, so secrets and access control belong in your own code, not in the instructions. Prompt injection, where text hidden in a web page, an email or a file tries to override the brief, is another reason for caution. Developers also use the term for the hidden instructions that apps add on top of a base model, which is why the same model can feel very different in two products. Brand voice guides, tutor personas, support scripts and code-review rules are common examples; engineers version them, run regression tests and log which revision answered. Providers usually place the system message at the top of the request and keep it constant across turns, which often lets them cache it, whereas user turns change every time; some models treat a developer message as a slightly different tier of instruction.

In the NeverNo app, custom instructions and characters set the role a model plays. You write how it should answer, or define a character with a personality and a manner of speaking, and the model keeps to it through the conversation. Your instructions and characters are stored in your browser and, if you turn sync on, follow you encrypted to your other devices.

In the API, a chat request accepts messages with the roles system or developer, alongside user, assistant and tool. NeverNo Uncensored, NeverNo’s own private model, is included in the Free plan with a context window of 128,000 tokens, and the page about uncensored AI chat shows how it answers frankly. A system prompt can shape its tone and format; the hard limits of NeverNo stay in place whatever it says: nothing involving minors, nothing about real people without their consent, and no real help to seriously harm others.