Context of Automation

A Site Reliability Engineer (SRE) recently discovered that he was spending 45 minutes writing a runbook for an alert he had already documented multiple times for other services. Although the structure of these documents was similar, the service name changed each time. Upon closer examination of how he was using his time, he realized that he was dedicating 3 hours per week to repetitive tasks such as writing runbooks from blank templates, generating standard Terraform code, manually constructing kubectl commands he had already typed hundreds of times, and drafting post-mortem documents while still exhausted from the incident. These tasks, while important, did not require critical judgment but simply time. This is precisely the kind of work for which Claude Code was designed.

Transforming Work Time

After integrating Claude Code into his processes, this engineer managed to transform those 3 hours of weekly work into just 20 minutes. Within six weeks, he modified his workflows, using specific prompts and a CLAUDE.md file to optimize the use of Claude Code. However, he identified two tasks that he still preferred to handle manually, highlighting the importance of human intervention in certain critical aspects.

What Claude Code Really Is (and Why It’s Different)

Unlike other AI coding tools that primarily focus on autocompletion in Integrated Development Environments (IDEs), Claude Code stands out for its ability to execute real commands in the terminal. It functions as an agent, capable of reading files, executing shell commands, modifying configurations, and interacting with tools like kubectl and Terraform. Claude Code uses the same SSH keys and cloud credentials already loaded in your environment, ensuring seamless integration. A key feature is its request for approval before each action, thus ensuring complete control over the operations executed. This approach guarantees that nothing runs without the explicit consent of the user, which is crucial for trust in infrastructure-related operations.

Automating SRE Tasks

The tasks that Claude Code helps automate go far beyond simple function writing. They include reading logs, building runbooks for alerts, generating Terraform modules compliant with existing templates, and drafting post-mortem documents based on incident timelines. Claude Code excels at managing complex tasks that require extensive context and direct integration into existing workflows. These tasks span multiple files and necessitate a deep understanding of the actual source code, producing results that integrate directly into already established processes.

Pricing (as of July 2026)

Claude Code offers several pricing tiers tailored to the varied needs of users. The Pro option is available for $20/month, including Claude Code and suitable for beginner users. For those needing enhanced capabilities, the Max 5x option is offered at $100/month, providing usage limits five times higher than the Pro option. Finally, for daily heavy use, the Max 20x option is available at $200/month. The mentioned engineer uses the Max 5x option, which proves cost-effective by saving 3 hours per week, far exceeding the break-even point of 2 hours per month.

Configuration: The CLAUDE.md File

To make the most of Claude Code, creating a CLAUDE.md file at the root of the repository is essential. This file provides the necessary context for Claude Code for each session, including team conventions, infrastructure templates, and SRE standards. Without this file, the results would be generic. With it, the results are tailored to the user's specific environment. This file, while taking 20 minutes to write, influences every output from Claude Code, ensuring that the results align with the organization's practices and standards.

Here is an example of a CLAUDE.md file for an SRE repository:

CLAUDE.md — SRE Infrastructure Repository

Context

This repository manages the SRE infrastructure for a multi-region AWS deployment, using technologies like EKS (Kubernetes 1.29), Terraform 1.8, Datadog for observability, PagerDuty for alerting, and GitHub Actions for CI/CD.

Coding Conventions

• Terraform modules are organized in /modules, and environments in /environments/{prod,staging,dev}.
• A remote state is always used, with an S3 backend and a DynamoDB locking table.
• Each resource must be tagged with information such as environment, team, service, and cost center.
• Kubernetes manifests must specify a namespace per service, with required resource requests and limits.

SRE Standards

• SLO targets aim for 99.9% availability for production services.
• Alert thresholds are triggered at 10% below the SLO, for example, a P99 latency above 450ms when the SLO is 500ms.
• Runbooks are stored in /runbooks/{service-name}/, named according to the format {alert-name}.md.
• Post-mortems follow a specific template located in /templates/postmortem.md.
• All kubectl commands must explicitly use namespaces, never the default namespace.

Incident Response

• Severity 1 incidents, which are visible to customers, require immediate alerting to the on-call team.
• Severity 2 incidents, which are degraded, must be notified to the team channel.
• Severity 3 incidents, which are internal only, result in the creation of a ticket.
• The runbook format follows a precise structure: Impact → Detection → Mitigation Options → Escalation.

What NOT to Do

• Never run terraform apply without explicit confirmation.
• Never delete resources without listing them first.
• Never commit credentials or secrets.
• Always validate Terraform with terraform validate before displaying output.

This file runs at the beginning of each Claude Code session, ensuring that all actions comply with established standards.

Workflow 1: Runbook Generation (45 min → 5 min)

Writing runbooks was a time-consuming task. Each new alert required a runbook, and while the structure was the same, the content varied. Now, with Claude Code, this process is simplified. By starting Claude Code at the root of the repository, the engineer can generate a runbook in just a few minutes. Claude Code uses the context from CLAUDE.md and checks existing runbooks for consistency, thus generating tailored and accurate documents. For example, for a PagerDuty alert regarding high memory usage of EKS nodes, Claude Code generates a detailed runbook with kubectl commands to diagnose the issue, identify common causes specific to the environment, and propose mitigation steps. This process, which initially took 45 minutes, is now reduced to 5 minutes, freeing up time for other critical tasks.