> ## Documentation Index > Fetch the complete documentation index at: https://docs.rootly.com/llms.txt > Use this file to discover all available pages before exploring further. # AI SRE Instructions > Add team-specific guidance that steers what Rootly AI SRE checks and how it reports every investigation. Instructions capture the investigation practices your team wants Rootly AI SRE to follow across alerts and incidents. Use them for durable guidance about evidence, system conventions, and report quality. Each Rootly team has one set of AI SRE instructions. It applies to every AI SRE investigation on that team's alerts and incidents. In the editor, "Applies to everyone on your team" means the currently selected Rootly team. ## Configure Instructions Paths on this page use the consolidated **AI SRE** navigation, which is rolling out. If your sidebar still shows **AI & Agents**, open the same cards under **AI & Agents → AI SRE**. See [Where AI Settings Live](/ai/ai-settings#where-ai-settings-live). Open **AI SRE → Atlas**, then select the **Instructions** card. The card shows **Configured** when instructions exist and **Not configured** when they don't. Select **Markdown** to edit. Describe the systems to check, evidence to prioritize, or investigation practices to follow. Use direct, testable language. Instructions are Markdown and can be up to 10,000 characters. Click outside the editor. There's no **Save** button: Rootly saves when the editor loses focus and shows **Successfully saved AI settings**. Changes apply to investigations from that point on. Select **Preview** to see the rendered Markdown. The page opens in **Preview** when instructions exist and in **Markdown** when they're empty. To remove the instructions, clear the editor and click outside it. Members with an Incident Response seat can view the instructions. Incident Response Owners or Admins, or On-Call Admins, who also have an Incident Response seat can edit them. Members who can view but not edit see the instructions read-only, with the footer **Read only — only admins on your team can change these instructions.** Members without an Incident Response seat can't open the page. See the [AI & Agents permissions matrix](/managing-users/user-permissions#ai-agents-and-ai-sre) for the full rules. Your Rootly account team turns on AI SRE for the selected team, separately from the **Opt in to Rootly AI capabilities** setting. The **Instructions** card and **Generate with AI** may not be enabled for every team, even when AI SRE is. Ask your Rootly account team to turn them on. See [Getting Started](/ai/ai-sre/getting-started). ## Review Version History After instructions have been saved at least once, editors can select **History** to open **Instruction History**. It lists the last 5 changes, each with the editor's name, whether the change **created**, **edited**, or **cleared** the instructions, when it happened, and a line-by-line diff. History stays available after you clear the instructions, and the next save continues the same history. ## Generate with AI If your instructions are empty, editors can select **Generate with AI** to draft a starting set. The button is disabled while instructions exist, with the tooltip **Clear the existing instructions before generating new ones with AI.** Generation runs in the background. Rootly shows **AI SRE instruction generation started**, and the button shows **Generating instructions…** until the draft is ready. To write the draft, Rootly reads: * Facts from your [Knowledge graph](/ai/atlas/knowledge-graph). * Your saved Rootly AI instructions and your investigation rules that aren't **Paused**, with their conditions and instructions. * Datadog environments, services, and monitors, and New Relic data schemas, when those [AI connectors](/ai/connectors/overview) are connected. * A small number of Notion pages that look like runbooks or on-call procedures, when Notion is connected. * The tool lists of your connected custom MCP connectors. The draft is Markdown with a short summary and up to eight sections of short rules, such as environment and query scope, telemetry query conventions, signal interpretation, change and deployment correlation, dependency investigation, runbooks, and evidence and action policy. It aims for stable rules about how to query and read your systems, not a snapshot of current topology or owners. When generation finishes, Rootly saves the draft straight in as your live AI SRE instructions and refreshes the page. The text starts with a heading that ends in (Draft), but it's live as soon as it's saved. There's no separate approval step. Investigations use the text as soon as it's saved, so review and edit it right away. **History** records the change under the name of the person who selected **Generate with AI**. Generation never overwrites instructions someone saved while it was running. If generation fails, the page shows no error. Rootly retries in the background. If the button still shows **Generating instructions…** after five minutes, reload the page. If the editor is still empty, select **Generate with AI** again. ## What to Include Useful instructions teach AI SRE something specific about your environment or evidence standards. Name the service aliases, environment labels, ownership boundaries, or topology details that aren't obvious from provider data. Identify the source of truth for a signal. For example: "Use the payment-api error-rate monitor to establish customer impact; don't infer impact from worker retry volume." State checks that should happen across every investigation. For example: "Compare the alert window with the previous 30 minutes and check deploys to the affected service and its direct dependencies." Define what a useful handoff contains. For example: "Separate confirmed facts from open hypotheses and name the missing evidence for every unresolved cause." ## Instruction Scope and Precedence Three kinds of guidance can reach an investigation. | Scope | Applies To | Behavior | | ----------------------------------- | ------------------------------------------------------ | ----------------------------------------------------------------------------------- | | **Global instructions** | Every supported Rootly AI feature on the selected team | General guidance | | **AI SRE instructions** | Every AI SRE alert and incident investigation | More specific than global instructions when they conflict | | **Investigation rule instructions** | Alert investigations that one matching rule applies to | Included with AI SRE instructions for that alert class, without a separate priority | Configure global instructions in the **Instructions** editor on **AI SRE → Atlas → Global** (**AI & Agents → Global** if your sidebar doesn't have an **AI SRE** item). If **Global** uses Rootly's **Best practice** instructions, that text also reaches investigations. See [AI Settings](/ai/ai-settings) for the team-wide AI configuration surface. AI SRE receives global instructions and a separate block of AI SRE guidance. Within that block, the rule's text appears before the AI SRE instructions, but this order does not give either one priority over the other. If global guidance conflicts with that block, the more specific AI SRE guidance takes precedence. At most one investigation rule adds its instructions to an alert investigation, and incident investigations never use rule instructions. See [Investigation Rules](/ai/ai-sre/investigation-rules) for which rule applies to automatic and manual runs. Instructions steer where AI SRE looks and how it reports. An instruction is never evidence: it can't prove a cause the evidence doesn't show. Instructions also don't grant permissions, expand connector access, or override Private Agent policy. The investigation still needs current evidence. ## Write Effective Instructions * Name the provider, service, signal, or environment that matters. * Ask for a comparison or verification that produces evidence. * Explain ambiguous naming, such as two provider labels that refer to the same service. * State how to handle uncertainty and what to include in the report. * Keep team-wide instructions broadly applicable. Put alert-specific guidance in an investigation rule. Avoid instructions that: * Assert a cause before the investigation starts. * Tell AI SRE to ignore contradictory evidence. * Prescribe a production change without a current safety check. * Depend on one incident's timestamp, identifier, or temporary state. * Repeat information already present in every alert payload. ## Examples ### Evidence Standard ```text theme={null} Treat the checkout service level objective (SLO) as the source of truth for customer impact. Corroborate it with payment-api errors and traces. If the sources disagree, report the disagreement instead of choosing one silently. ``` ### Deployment Correlation ```text theme={null} For a service regression, check deploys to the affected service and its direct dependencies during the 30 minutes before the signal changed. A nearby deploy is a candidate cause until logs, metrics, or traces show the mechanism. ``` ### Report Quality ```text theme={null} Separate confirmed facts, likely causes, and open questions. For every open question, name the evidence that would settle it. ``` ## Related Pages How AI SRE investigates, the outcomes it reports, and its settings. How AI SRE uses instructions without treating them as evidence. Apply guidance to a specific class of alerts and choose when it runs automatically. Review the evidence-backed learnings AI SRE can recall when they're relevant. See what an investigation reads and where its access stops. Configure global instructions and other Rootly AI settings. ## Frequently Asked Questions No. Instructions tell AI SRE what to check, while an AI connector or Private Agent supplies the evidence. If the source isn't connected or AI SRE isn't permitted to reach it, the investigation can't draw on it. Put a short, broadly applicable investigation practice in instructions. Memory holds learnings AI SRE proposes from past investigations. Admins review them, and a note is used only once it's active. You can't write a Memory note yourself. See [Memory](/ai/atlas/memory). Connect the full runbook through a knowledge provider such as Notion, Confluence, or a custom Model Context Protocol (MCP) server. Both pieces of AI SRE guidance are supplied to the investigation, with no set priority between them. Avoid conflicting wording by keeping team-wide instructions general and rules specific to their alert class. No. Each Rootly team has one set. Use [investigation rules](/ai/ai-sre/investigation-rules) for guidance that should apply only to certain alerts.