System Prompt Conflicts

What happens when your system prompt says one thing, but your user prompt says another? This is a conflict, and understanding how the AI resolves it helps you write better prompts.

Imagine this system prompt:

Always respond in formal English. Never use contractions.

Then you send this user prompt:

Reply casually, like you're texting a friend. What's the weather like today?

Now the AI is stuck.

The system prompt demands formality. The user prompt demands casual language. Which wins?

Generally, the system prompt takes priority, but not always absolutely. How firmly it holds depends on the model: large models usually side with the system prompt, while small models often follow whichever instruction came last. The AI tries to balance both, which can produce awkward, inconsistent results. It might respond formally but attempt a friendly tone, creating a strange hybrid.

Conflicts also happen within a system prompt. Consider:

Keep responses under 50 words. Always provide detailed, thorough explanations.

These rules contradict each other. The AI can't be both brief and thorough, so it compromises, often satisfying neither requirement well.

To avoid conflicts, review your system prompt before adding new rules. Ask yourself: do any of these instructions pull in opposite directions? If you want the AI to follow a rule consistently, make sure nothing else undermines it. Clear, non-contradictory instructions produce reliable behavior.