Prompts
A conversation never starts empty. Before your first word the model has already been told today’s date, when to look something up, what to do with search results, how to write a summary if the conversation gets long. That text is not hidden, and it is not fixed. It all lives in Settings → Prompts.
The tab holds two kinds of instruction, and the difference matters.
The system prompt
Section titled “The system prompt”What you want the model to be like: a tone, a language, a rule it should always follow. There is a global one, and you can give a specific model its own, either extending the global one or replacing it.
This is the prompt most people mean when they say “system prompt”. It is yours, it is short, and nothing in the app writes to it.
What the app adds on its own behalf
Section titled “What the app adds on its own behalf”Everything else on the tab, grouped by the part of a turn it governs:
| Group | What it decides |
|---|---|
| Conversation | The date line, several participants, quick-choice questions, the title |
| Personas | The language rule, being summoned with @, an exchange brought back |
| Web search and pages | Whether to search, the query, how results and pages are used and cited |
| Memory | What a persona keeps about its user, and the tools that change it |
| Images | Writing a prompt an image model can draw, and naming what comes back |
| Native tools | The same instructions for providers that call tools instead of reading |
| Compaction | Writing a summary, using one, and anything typed after /compact |
Each prompt is folded, showing its name and what it drives. Open it to read the wording, edit it in place, and it saves as you type. Blank it, or type the default back, and your version disappears.
Some take placeholders, listed under the editor: {datetime}, {results}, {pages}, {language},
{summary} and a few more. They are substituted at send time. Leave one out and the model simply
never sees that part.
Why the native tools say it twice
Section titled “Why the native tools say it twice”There are two ways to give a model the web. Most endpoints get a text protocol they answer in prose; providers that support tool calling get the same instructions as structured tool descriptions. Both are on the tab, next to each other, because they have to agree. If they drift, the app behaves differently depending on which provider answered, which is the kind of bug nobody can reproduce.
What is not here
Section titled “What is not here”A persona’s prompt and a playbook’s instructions. Those belong to something you wrote and stay with it, in the library. The tab is for what the instance says, not for what you built.
On a shared instance
Section titled “On a shared instance”An admin can share both kinds of prompt, locked or as a starting point. Everyone still sees the whole tab either way: a read-only prompt tells you what is being sent on your behalf, where a hidden one only tells you that something is.
Shared as a starting point, the rewrites merge prompt by prompt. Rewriting the summary keeps the admin’s version of every other one, because twenty prompts are twenty settings that happen to share a screen.
Locked is enforced when a turn starts, not only in the interface. See Administration.