Overview
Content Manager
Categories
AdSense
Settings
View site
Edit post
1 min read · published
Live preview
Save post
Title
Slug
/blog/prompt-engineering-for-ops
Excerpt
Treat the model like a junior engineer with perfect recall and zero context. Structured prompts, guardrails, and where automation should stop.
Content (Markdown)
## Context is the whole job A model cannot infer your naming conventions, your firewall policy or the fact that the staging box shares a database with production. Everything it does not know, it invents. ## A prompt template that holds up ```text ROLE: senior Linux administrator CONTEXT: Ubuntu 24.04, nginx, PHP-FPM 8.3, no Docker TASK: write an idempotent bash script that rotates logs older than 14 days CONSTRAINTS: POSIX-safe, no third-party binaries, dry-run flag required OUTPUT: single script, comments only where non-obvious ``` ## Guardrails before autonomy Every generated script gets a dry-run mode and runs against staging first. Automation that cannot be previewed is not automation, it is gambling. ## Where to stop Do not automate DNS cutovers, billing changes or anything with an irreversible blast radius. The time saved never covers the one bad night.