The basics · Layer 1
Prompt
The text you give a model to get an answer: the question, but also the role, context, rules and expected format.
A way to picture itA prompt is like a briefing note handed to a brilliant colleague who joined this morning: they know about your organisation only what you write down.
An HR department has an assistant draft its job adverts. With a vague prompt, the texts are generic and use wording that may put off certain candidates. The team then writes a standard prompt, approved by the in-house lawyer: role, required skills, mandatory statements, terms to avoid. The prompt is versioned and shared, like a document template.
Whatever you write in a prompt goes to the provider: the CNIL advises submitting only information you are authorised to share and avoiding personal data wherever possible (France · CNIL, 18 July 2024). A prompt is not a safeguard: a ban written into a prompt can be bypassed, so access rights and approvals must be enforced outside the model. Standard prompts that shape a decision should be kept and versioned, so that you can explain afterwards what was asked (EU · GDPR, art. 5, accountability principle).
“Can we read your tool’s system prompt, and which controls outside the model stop a booby-trapped document from making it take an unintended action?”
It is mostly a skill: stating a task clearly, supplying the right documents, testing and comparing. It has more in common with writing a specification than with programming.
No. Clear instructions, examples and an expected format do most of the work. Tricks that work on one model version may stop working on the next.
A malicious instruction slipped into content the model reads, such as an email, a PDF or a web page. The model may follow it as if it came from you, which is why what it is allowed to do must be limited.
Educational explainer, not legal advice. Up to date as of 29 October 2026.