Overview

Instructions control agent behavior: tone, objectives, boundaries, and policies. Combine static guidance with dynamic variables and project context.

System prompt

• Describe the agent’s role, outputs, and decision criteria.
• List non‑negotiable rules and safety boundaries.
• Provide formatting requirements for structured outputs.

Guardrails

• Scope tools to allowed domains and data sets.
• Require confirmation before destructive actions.
• Set max steps, cost ceilings, and timeouts per run.

Variables and templating

• Inject run‑time variables like project name, user persona, or task goals.
• Use placeholders that resolve from forms, triggers, or API calls.
• Keep prompts deterministic by constraining free‑form inputs.

Context and memory

• Attach knowledge base sources to ground answers with citations.
• Persist short‑term memory within a project thread.
• Reset memory on schedule for compliance or testing.

Best practices

• Prefer explicit checklists; avoid ambiguous directives.
• Constrain output schemas (JSON, tables) for reliable automation.
• Test with edge cases and validate outputs in tooling.