At ThrillTech, we are deeply invested in understanding and using AI technology as a tool. As with all things, we always advocate using the right tools at the right time. AI definitely has its place in everyone’s lives, but at the engineering level, we have learned (and experienced first hand) how AI can both accelerate and completely obliterate the productivity of a team due to subtle hallucinations, poor training data and the now-infamous confidence and swagger that AI always seems to deliver, no matter how invalid its assertions are.
With our collective experiences in hand, we decided to create our 10 Commandments for AI use in Engineering (read: points of guidance) for everyone in our engineering team. While the commandments themselves have been written by hand, all the images were AI generated (obviously!).
We are sharing these with you, the reader, since we thought they may be both entertaining and potentially impactful on how AI is deployed in other engineering teams.
1. Thou Shalt Not Treat AI as an Oracle
AI is not the All-Knowing Algorithmic Overlord(TM).
It’s more like a slightly over-caffeinated junior dev who’s read all of Stack Overflow, but still occasionally confuses Java with JavaScript.
2. Trust, but Also Verify (and Then Double-Check the Specs)
AI can draft the function, suggest the API call, or recommend the library —
but it’s still your job to make sure it won’t crash in prod or rm -rf / the wrong directory.
3. Keep the Human in the Loop (Even if the Loop is Chaotic)
Your job is to review, refactor, and steer AI output.
Remember: AI doesn’t know if you’re building a login form, a payment system, or the next great infinite loop unless you tell it.
4. Feed It Data as You Would Feed a Production Database
Garbage In = Garbage Out.
Or worse, Well-Formatted Garbage Out — the kind that comes with neat JSON, unit tests, and a reassuringly green CI pipeline… right before it corrupts your schema.
5. Mind the Bias — It’s Sneaky
AI can inherit biases from its training data,
just like new devs pick up “standard practices” from seniors who once fixed a production outage with a hard-coded sleep(10).
6. Document the Madness
When you use AI for decisions, log the prompts, versions, and reasoning.
Future You will thank Past You when debugging at 3 a.m.
(and wondering who thought // TODO: fix later was acceptable).
7. Don’t Let AI Code Blindly in Production
“Ship It” without review is how you end up in the Incident Retrospective Hall of Fame, with your story forever remembered as The Deployment That Brought Down the Cluster.
8. Respect IP and Confidentiality (Or the Legal Team Will Find You)
Don’t paste your API keys, client data, or the secret YAML that actually keeps prod running into the prompt box.
AI is a tool, not a password manager.
9. Beware of AI’s Confident Wrongness
AI doesn’t hesitate — it will suggest broken regex with the swagger of a 10x engineer.
That’s why you’re here: to spot the nonsense before it hits version control.
10. AI is a Power Tool, Not a Replacement Engineer
Treat it like an IDE plugin on steroids — it can autocomplete half your brain,
but you still write the architecture (and stop it from eval()ing user input).
