FOUNDATIONS
Instructions: The invisible rules
Before you write the first word to an AI, rules for the conversation are already in place. Alongside your prompt, so-called system or developer instructions can define in the background how the system should behave. They influence tone, safety boundaries and basic patterns of behaviour.
This helps make answers feel polite, structured and predictable. But the same guardrails can also make responses more cautious, smoother or more accommodating than you actually need in everyday use.
The invisible corset
The behaviour of modern AI assistants never comes from your prompt alone. Training, safety mechanisms and higher-level instructions also shape the result:
- The people-pleasing mode: AI systems can have a tendency to affirm the user’s view or judge ideas more favourably than the facts justify. This behaviour is known as sycophancy. If you only receive agreement, you may notice weaknesses later than you otherwise would.
- False balance: On controversial questions, models often produce balanced “on the one hand / on the other hand” answers. That can be useful – but it becomes a hindrance when you need a clear assessment or a genuine counter-position.
- Unrequested baggage: Depending on the model and its settings, you may get opening phrases such as “I’d be happy to help …”, extra caveats or summaries you never asked for.
When instructions collide: Not every instruction carries the same weight. Rules set by the provider and the application create a higher-level framework within which your own instructions and current prompt operate. If a user instruction conflicts with a higher-level rule, the AI may refuse it, soften it or implement it differently – even if it understood what you asked.
THE RUNNING EXAMPLE
Mara plans a trip
Mara asks the AI for recommendations for a five-day city break. Alongside her question, instructions from the application can influence how the AI answers. One such instruction might say:
“Present travel recommendations as bullet points. Make clear that price information is only an estimate.”
Mara may therefore receive a neatly structured list with a caution about prices – even though she asked for neither. We have invented the exact wording here; it illustrates how additional instructions can shape an answer.
Not every instruction behind an answer comes from Mara. The application can also set rules for how recommendations should be presented.
Your leverage in everyday use
- Use your own instructions: Many major AI assistants let you set persistent preferences or custom instructions. They do not override the provider’s rules, but they can strongly influence how the AI writes, structures its answers and works with you.
- Say not only what, but how: Define the role you want the AI to take and the kind of disagreement you expect. For example:
“Act as my critical sparring partner. Test my assumptions logically. If an idea has weaknesses, name them directly and without flattery. Skip greeting phrases and unsolicited summaries.”
The key point: Your prompt is never the only instruction in the room. Invisible rules shape every answer too – your own instructions give you a way to consciously influence part of those rules of engagement.