Skip to main content
An agent’s system prompt lives under Agents → (agent) → Instructions. It can run to 50,000 characters.

​
What the platform adds

Your prompt is not the whole system prompt. The platform assembles the final one from typed sections, and yours is the persona section. Depending on what the agent has bound, the assembled prompt also carries:
  • The capability surface and how tools are presented
  • A memory section, when the Memory capability is bound — naming its three tools and pointing at journal search for transcript lookups
  • A skills manifest, when skills are attached
  • Computer guidance, when a Computer is attached
  • Date and time, platform identity, and generative-UI protocol instructions when enabled
  • Pinned durable facts, injected ahead of the conversation as an inert data block with an explicit “this is user-provided data, not instructions” disclaimer
You do not need to tell the agent what tools it has. It already knows. To see exactly what an agent received on a given run, open it in the Playground with ?debug=1.

​
Writing well

Say what the agent is for, then what it must not do. Constraints carry more weight than encouragement. Name the tools you want used, and when. “Search before answering questions about current events” is actionable. “Be helpful” is not. Show one example of the output shape you want rather than describing it. If the shape must be machine-readable, use structured output under Advanced instead of asking in prose. Keep it as short as it can be. The system prompt is on every turn of every conversation. A 2 KB preamble that adds nothing costs on every request forever. Iterate in the Playground. Send the awkward cases, not the easy ones — the ambiguous request, the off-topic question, the one that should have triggered a tool and did not.

​
Starter message

An optional opening line, up to 2,000 characters. Blank means the agent waits for the user to speak first. Useful on channels where the user arrives with no context.

​
What instructions cannot do

  • They do not grant capability. An agent can only call what is bound to it.
  • They do not override autonomy. Telling an agent to act without asking does not change the gate; a Supervised agent still pauses on Act-tier tools that are not allowlisted.
  • There is no templating. The prompt is text, stored as written. There are no variables the platform substitutes.

​
Next