Last month I had ChatGPT reply to a message from a difficult client. The AI sounded polite and reasonable. The client blocked me.
That near-miss sent me to a Google veteran
I thought I was fine with AI until last Wednesday, 11pm. My partner and I were tweaking a landing page in Cursor. The AI spit out 80 lines of code in one go. We didn't even glance at it before pushing live. Next morning the client called: "The button doesn't respond when I click it."
That stung, so I went and listened to the Pragmatic Engineer podcast. The guest was Addy Osmani — 14 years at Google, led teams behind Chrome itself and the DevTools panel that pops up when you hit F12. He's been steering AI developer experience lately.
Addy dropped a term that hit me hard: loop engineering. In plain words: "AI finishing the work doesn't mean the work is done. You have to look it over again."
His own example was my exact mess
He said he'd asked AI to help draft a technical doc the week before. The AI's answer looked airtight. He spent another 20 minutes going line by line and found three spots where the AI had invented API usage out of thin air (APIs are how software talks to other software) — if that doc had gone out, every teammate using it would have hit a wall.
He called the mindset behind that cognitive surrender: you trust the AI's answer so much you stop thinking for yourself. I broke out in a cold sweat — wasn't that exactly what happened with my client reply?
Replicating this playbook today, $0
Money: $0 — the ChatGPT, Claude, or Cursor you already have works.
Time: +5-10 minutes per task.
Technical barrier: can copy-paste, can read with your eyes.
First step: today, before you let AI draft anything that goes out the door, treat it like "an intern's first submission" — read it cover to cover. Circle any word you're unsure about and ask the AI: "How do you know this?" If it can't point to a source, it's making it up.
Which stage are you at?
If you're just starting out (one-person company, 0-10 clients): don't let AI reply to messages for you yet. The high-stakes moments — pricing, negotiation, apologies — write those yourself. AI is for drafts only.
If you have 1-2 long-term clients: you can let AI draft client emails, but you need the "second pass" wired into your head. Read it once yourself, strip out the "Dear valued customer" AI flavor, then send.
If you're scaling (you've hired help): put loop engineering straight into your SOP. Anyone on the team sending anything external has to tick a checkbox that says "I read this at least once." That's how Addy ran things at Google.