# System prompt and revisions
> Write the standing instructions an agent reads before every conversation, keep every version, and roll back when a change makes the agent worse.
The **system prompt** is the standing instruction an agent reads before every conversation: what it does, how it should decide, what it must never do. Every version you save is kept, so changing it is reversible.
## When to use it
Reach for the prompt when the agent's *behavior* is wrong — it does the wrong thing, in the wrong format, or asks you things it should already know. Reach for a [connector](/docs/connect/overview/) or a [skill](/docs/capabilities/skills/) when it lacks a *capability*. Prompting cannot substitute for access.
## Write one
1. **Open the prompt section in [Agent Builder](/docs/build/agent-builder/).**
You see the current text and the history of previous revisions.
2. **State the job in one paragraph.**
Who the agent works for, what it is responsible for, and what falls outside that. An agent that knows its boundaries asks fewer pointless questions.
3. **Say what a good result looks like.**
This is the part most prompts skip, and the part that changes results most. The agent chooses its own steps; what it cannot guess is your standard for "done".
*"Summarize my unread email"* leaves everything to chance. *"Summarize my unread email, grouped by sender; for anything asking me a direct question, quote the question verbatim so I can answer without opening it"* does not.
4. **List the hard rules.**
Things that must always or never happen — never email anyone outside the company, always cite the source document, never guess at a number.
5. **Save it as a new revision.**
The previous text is not overwritten; it becomes a revision you can return to.
## Revisions and rollback
Every save creates a revision. The history shows what the prompt was at each point, and you can restore an earlier one.
This is what makes prompt iteration safe rather than a gamble. Change something, use the agent for a day, and if the change made it worse, go back to the revision that worked — you do not have to remember what the old text said.
## What belongs where
| Put it in | When |
| --- | --- |
| **System prompt** | It applies to every conversation — your role, standing rules, output format |
| **A message** | It applies to this task only |
| **[Character](/docs/build/identity/)** | It is about tone, not about the job |
| **[Memory](/docs/capabilities/memory/)** | It is a fact the agent should learn once and reuse — a preference, a name, a habit |
| **A [skill](/docs/capabilities/skills/)** | It is a repeatable procedure with steps, not a standing instruction |
A prompt that has grown to cover four unrelated jobs is a sign you need a second agent, not a longer prompt.
## Limits
- A prompt cannot grant access. If the agent needs your calendar, it needs a bound connection — writing "you have access to my calendar" changes nothing.
- Restoring a revision replaces the whole prompt. There is no merge between revisions.
- The prompt is per agent. Two agents that should share standing rules need the text in both.
## Related