Declarative, per-integration guidance that tells the agent when to call this integration and how to use it safely. Keep it outcome‑focused, scope‑limited, and explicit about side effects.
Agents discover tools dynamically. Clear instructions reduce hallucinated calls, prevent unsafe writes, and keep cost/performance predictable.
Be declarative: describe what to accomplish and when to use this integration, not step-by-step mechanics.
Constrain scope: enumerate allowed capabilities and explicitly forbid risky actions.
Name preconditions: required fields, auth, scopes, and object existence checks.
Define side effects: what gets written/updated; link to approvals if destructive.
Last updated 5 months ago
