Configure Assistant rules

Custom rules guide how Grafana Assistant responds and behaves. Rules are automatically applied to every conversation and can target specific applications. Due to the indeterministic nature of LLMs, there’s no guarantee they are followed exactly as specified, but they strongly influence the Assistant’s output and decision-making.

You can create rules in Settings > Custom rules and apply them to specific scopes:

Just me: Rules apply only to your own conversations.
Everybody: Rules apply to everyone in your stack (requires admin permissions).

Change review applies Everybody rules for the Assistant application to the pull request reviews it posts on GitHub. It never applies Just me rules, because a review runs for your whole stack.

Create a rule

Navigate to Assistant > Settings > Custom rules.
Click Create rule.
Enter a rule Name and content.
Choose the Scope: Just me or Everybody.
Select the Applications the rule applies to, such as Assistant, Pages, Investigations, or Infrastructure memories. The Pages option appears when Grafana Pages is enabled in your environment. Grafana Pages is in private preview.
Click Create rule.

The Assistant applies your rule automatically in new conversations. Rules that target Pages load when starting or resuming a Grafana Pages sandbox and apply specifically to creating or editing pages rather than general Assistant chat.

Rule examples

Use these examples as starting points for your own rules.

Grafana Pages guidelines

Guide how the Assistant builds and formats Grafana Pages:

When creating or editing Grafana Pages, use two-column layouts for dashboard panels, add explanatory markdown callouts above key charts, and ensure all visualization panels define clear units.

Infra and app hints

Guide the Assistant to use specifics relevant to your environment:

When discussing CPU metrics, use container_cpu_usage_seconds_total instead of generic CPU references. Reference our custom service_health_score metric for overall service status. Prefer the prometheus_prod datasource when investigating issues in production.

Communication style

Define how the Assistant formats and structures responses:

Always be concise and direct in responses. Use bullet points for actionable recommendations. Avoid jargon unless specifically discussing technical implementation.

Best practices

Encode organizational standards and observability methodologies:

Use the RED method (Rate, Errors, Duration) wherever possible. Always recommend setting up SLIs before creating SLOs. Suggest using template variables for dynamic dashboards.

Workflow guidelines

Automate operational recommendations based on specific triggers:

Suggest declaring an incident if the ecommerce platform response time exceeds 5 seconds. Recommend escalating to the infrastructure team for persistent memory issues. Always ask about recent deployments when troubleshooting performance issues.

While Rules provide general behavioral context, other features offer more structured guidance:

Skills: Create step-by-step troubleshooting guides for specific alerts or scenarios. Skills are triggered by specific events, whereas Rules apply generally to conversations.
MCP Servers: Extend the Assistant’s capabilities by connecting to external tools and data sources.