The instructions you never see
Every assistant runs with a hidden block of instructions placed before your first message. It usually sets a persona, a caution level, formatting rules, and a list of things to avoid. The same model behind 2 different system prompts produces answers that feel like 2 different products.
People compare assistants and conclude that one model is more direct than another. Often the models are similar and the instructions differ, and the instructions are the part nobody publishes.
On this page
What the block usually contains
A persona sentence, a caution instruction, formatting preferences, and a list of subjects to handle carefully. It is written in ordinary language rather than as code, which is why its effect is felt as personality. A single sentence telling the assistant to be helpful and harmless changes thousands of answers.
Why it beats the model for tone
Tuning shifts a model's tendencies; the system prompt applies on every request on top of them. When the 2 disagree, the instruction in front of the conversation usually wins, because it is the most recent and most specific thing the model was told. This is why the same open model feels cautious in one product and blunt in another.
Length and hedging are often set here
Instructions such as be thorough or acknowledge uncertainty produce the padding and the qualifiers people blame on the model. Neither instruction is unreasonable on its own. Together they produce the long, careful, uncommitted answer that is the default experience of most assistants.
What you can tell from the outside
You cannot read the block, and you can measure it. Ask for a 1-sentence answer: a service that still produces 3 paragraphs has a length instruction that overrides yours. Ask the same factual question on 2 services: matching facts with different tone points at the instructions rather than at the models.
Reasonable questions
Can I override the system prompt?
Partly and not entirely. Formatting requests usually pass, and caution instructions usually hold, because they are placed as constraints rather than as preferences.
Do services publish their system prompt?
Very few do. It is treated as product design, and it is the single piece of information that would explain most of the differences between assistants.
Is a shorter system prompt better?
Shorter usually means fewer competing instructions, which produces more predictable answers. Longer is not automatically worse; contradictory is.
Ask for a 1-sentence answer here and count the sentences.
Open unfiltered chat