I made the deepest cut into my finger in my life. I put pressure on it and started a post mortem as I waited for the bleeding to stop.

My tool was sharp. I had gripped it the way I was taught at a knife skills class a week earlier.

I was cutting into a slightly older, rubbery red pepper that slipped at the last moment. I did not adjust my technique for the ever-changing real world.

Every time I stand at that cutting board, there’s about a 2% chance that I nick myself. My risk of injury goes up the longer it has been since I applied the last band-aid.

I still cook because the payoff of a good meal outweighs the chance of a minor injury. Each time I cut myself, the shame of the cut hurts more than the pain.

As an AI Chef, I choose the best tools and sharpest models for the job. I structure my prompts carefully.

Every single time I run a quality check on an AI output, I catch a mistake. Not sometimes. EVERY TIME.

It’s not because AI isn’t good enough. It’s because I ask it to do hard things. The most common mistakes I see across all the major systems: not following all the instructions in a prompt to the letter, or being technically accurate while missing the bigger strategic picture.

In a 2013 TED Talk, Bill Gates said: “Everyone needs a coach… We all need people who give us feedback. That’s how we improve.”

A top athlete can’t spot their own flaws mid-swing. AI can’t find its own areas of improvement while it’s focused on execution. The discipline is learning to coach it into top performance by talking in a language it understands.

I’ve found the best results from a two-step method.

First, I ask the assistant to perform a mechanical check against all the instructions and feedback given throughout the conversation. It works like a vacuum, systematically picking up debris.

Next, I ask it to look at the finished work through the lens of my original goals and suggest improvements that will make the biggest impact. It works like a broom, sweeping up the dust the vacuum missed.

I return to the cutting board more carefully, knowing that there’s still a 2% chance I’ll cut myself. It’s going to continue to happen (sorry Mom) because I’m sometimes careless and impatient.

I’m just as impulsive at the keyboard. I want to get the work done quickly and move on. But after catching something every single time I run these steps, I knew I couldn’t bear the shame of sending out work that wasn’t my best. The QC step is how I close that gap.

How do you check your AI work before it goes out the door?

Originally published on LinkedIn