Integration instructions

Guidance for when/how to write additional instructions for integrations

Persistent guidance

Integration specific agent instructions are included in every AI run that uses that integration. Use it to fill in context the AI cannot discover on its own, or to correct behavior you've observed going wrong consistently.

When to use

Only add instructions if you've experienced specific undesired behavior or the AI is missing context it consistently needs. Leave blank if things work as expected.

What to put here

• Domain terminology or entity relationships unique to this system (e.g. "user" in this API means a tenant; tenants belong to units, units belong to projects)

• Persistent defaults you always want applied regardless of the task (e.g. "default ticket status on creation is new")

• Account-specific structure the AI cannot discover from the API alone

What not to put here

• Task-specific instructions — those belong in the run intent or context

• Procedural steps for how to do things, unless that exact behavior is needed in every run without exception. Because these instructions are always present, procedure that conflicts with the current task or triggers unnecessary actions leads to worse results. Prefer describing what things are over how to handle them

Format

Plain text or simple markdown. One bullet or sentence per rule. Shorter is better.

For managed integrations

Global guidelines for this integration type are already present and applied automatically — you don't see them and your instructions don't replace them. What you write here is added on top, so focus on what is specific to your account or connection.