AI-assisted engineering cheatsheet
A one-page reference for using AI assistants safely in engineering/SRE work. For rollout strategy and productivity data, see the complete guide.
📖 Full guide: AI-Assisted Engineering →Never blindly apply AI-generated infra changes
- Treat AI-generated Terraform/K8s/IAM changes like a junior engineer's PR —
read every line, run
terraform plan/kubectl diff. - Be extra suspicious of AI-generated IAM/permission changes — models default toward overly permissive policies. Scope them down explicitly.
- Never let an agent apply to production without a human-approved plan-review-apply gate.
Human-in-the-loop, by risk
| Task | Oversight needed |
|---|---|
| Read-only (query logs, summarize metrics) | low |
Anything that writes (deploy, config push, migration, kubectl apply) | explicit approval gate |
Verify, don't trust
LLMs hallucinate plausible-sounding but nonexistent flags/config keys/API methods, especially for less-common tools or anything past the model's knowledge cutoff. Check generated commands against real docs or a dry run before running them against anything that matters.
Prompt engineering basics
- Give it a role/context to calibrate tone/expertise.
- Show 1-2 examples for a specific format.
- Ask for step-by-step reasoning on hard problems.
- Specify format explicitly (length, structure).
Overreliance risks
- Accepting generated code/config without understanding it — you own it in the incident at 3am, not the model.
- Skill atrophy on fundamentals the assistant now does by default.
- False confidence from fluent-sounding but wrong output.
Rollout checklist
- Start with low-risk, read-only use cases.
- Require human approval on anything that writes to production.
- Review flagged/high-risk outputs as a team, not solo.