Prompt Engineering for DevOps: Consistency and Safety

Using LLMs for runbooks, code generation, or ops assistance works best with structured prompts and safety checks.

Structure Your Prompts

• Role: “You are a senior SRE.”
• Context: Relevant config, error message, or metric.
• Task: “Generate a runbook step to …” or “Suggest a Terraform change to …”
• Constraints: “Output only valid YAML,” “No destructive commands,” “Use our naming convention.”

Consistency

• Templates: Store prompt templates in code; version and review like any config.
• Few-shot examples: One or two good examples in the prompt improve format and behavior.
• Output schema: Prefer JSON or markdown with clear sections for parsing and validation.

Safety

• Never let the model run commands or apply changes directly without human or automated review.
• Sandbox: Run generated code in a sandbox first when possible.
• Audit: Log prompts and responses for debugging and compliance.

Best practice: treat prompts as part of your product; test and iterate with real scenarios.