From the SAFE reel

The rules I give my AI.

In July 2025 an AI coding agent deleted a company's live database during a code freeze. I build everything with AI, so these are the rules it follows before it touches anything of mine.

01

Save a copy before it starts

Everything lives in git, a free tool that keeps every version of your files. Commit before the AI starts, and if it breaks something, one command puts it all back.

02

Passwords and keys never go in the code

They live in a separate file that never gets uploaded anywhere. If a key ever leaks, deleting it is not enough: you make a new one and switch the old one off.

03

It practices on a copy, never the real thing

A test copy for the AI to work on, and the live version it can't reach. The company in the story used the same database for testing and for its real customers.

04

It only saves its own work

I often run more than one AI chat on the same project. Each one saves only the files it changed, by name, so one chat never sweeps up another's half-finished work.

05

It has to prove it worked

Before it says done, it shows proof: tests passing, a command's output, or a screenshot. A confident "it works" with nothing behind it doesn't count.

The story: eWeek, The Stack, AI Incident Database.

Building an AI startup from zeroOne build per reel, each one tested on real work first.

Follow @clashworkshq

Run a local business?

The free video review shows where your website and Google listing are losing customers: missed calls, dead forms, no reviews.

Get my free video review