Instructions are standing guidance you write for an agent. They shape how the agent responds, what it prioritizes, and what it avoids, every time it’s tagged in a thread.
Think of instructions as a standing brief: the context a good employee would have read before they started.
Where instructions live
Global instructions apply to the agent everywhere, in every Space it’s added to.
Space-specific instructions apply to the agent only in a particular Space, and override the global instructions where they conflict.
What to include in instructions
Good instructions cover:
- Role: what this agent is responsible for in this context
- Tone: how it should sound (formal, direct, casual, technical)
- Format preferences: bullet points, prose, tables, short or long responses
- What to prioritize: which tools, which decisions, which context to lean on
- What to avoid: topics outside its scope, assumptions it shouldn’t make
- Recurring context: information it should always have in mind
Example instructions
For a marketing agent in a specific client Space:
For a support agent:
Tips for writing good instructions
- Be specific. “Be helpful” is not an instruction. “Respond in bullet points and keep each point under 15 words” is.
- State constraints clearly. If there are topics or actions outside scope, say so explicitly.
- Update them when the work changes. Stale instructions lead to misaligned responses.
- Test with real questions. Write an instruction, tag the agent with a real question, and see if the output matches your expectations.
Write instructions as if you’re onboarding a new team member to a specific role. What would they need to know on day one?
Character limit
Instructions are capped at 2,000 characters per instruction set (global and per-Space). Keep them tight. If you find yourself writing more, split the context into pinned docs the agent can read instead.
