Someone at the weekly AI Mastermind I host said it almost as a joke: “Don’t be a meat proxy.” A few people laughed. I wrote it down and sat with it for the rest of the call, because I wasn’t sure the joke was on someone else.
A software engineer named Niklas Gruhn wrote about it in August, and the term spread fast across programmer forums.
A meat proxy is a person who takes what an AI tool produced and passes it along without checking it, editing it, or understanding it. The human is just the cable between the chatbot and the next person.
Here is why that matters to you, even if you never open an AI tool yourself.
Someone on your payroll, or a vendor you pay every month, is probably doing it right now. The email your customer got this morning. The proposal that went out Friday. The website update your developer pushed last week. Your name is on all of it, and you may have no way of knowing whether a person actually thought about any of it.
So I asked myself the uncomfortable question first. Am I a meat proxy?
I use Claude every day. I build software for clients, and I am not a trained software engineer. Do I copy and paste some of what it gives me? Yes. If copying were the crime, I would be guilty.
But copying isn’t the crime. A carpenter with a nail gun isn’t faking carpentry. The question is whether your judgment is in the loop, or just your hands.
So I wrote four questions and answered them as honestly as I could. I call it the Proxy Test.
Could I explain it without the screen? I only build what I can understand, and I write down every decision as a ticket (a short written job description) on my project board before anything gets built. If a client asks why something works the way it does, the answer is in my head and on the board, not buried in a chat window.
Did I decide what mattered? AI can produce a screen in seconds. It does not know the difference between a screen that works and one that is good. I redesign a lot of what it hands me, because the gap between good and great is still a human call.
Would I sign it? My name is on everything I deliver. In my line of work, you are only as good as your last project.
Who catches the mistake? Me, at first. And because I know I will miss things, every app I build has a Feedback tab. When a customer reports a problem during beta testing, it captures what they were looking at, screenshot included, and drops it into a hub where I review and sort it.
Every professional needs a backstop. A meat proxy skips it, because they don’t think there’s anything to catch.
By my own test, I mostly passed. Then, this week, I found a hole in my own work.
I am building a member app for an outdoor recreation club. Members sign in, check trail conditions, let others know they are riding today, and decide who can see their phone number and email. Claude designed and built the screens where members manage those preferences. They looked good. Clean, clear, easy to use.
On a phone, you could sign out. On a desktop, you couldn’t. There was no way out.
I found it. Then I put a screenshot in front of Claude and asked it to look hard. It found three more. The header was still inviting an active member to “Join the Club.” The footer was still offering “Member Sign In” to someone already signed in. And a menu bar that sticks to the top of the screen was covering up a heading.
Three of those four were the same mistake. The page didn’t know you were signed in. Nobody told it the header and footer should change once someone logs in. I didn’t write it into the ticket, and Claude didn’t ask.
AI builds what you asked for. It doesn’t build what you assumed.
That is the line between a meat proxy and a builder. A meat proxy ships what came back. A builder holds what came back up against what they assumed, and goes looking for the gap.
Notice what found the other three, too. The same AI that missed them caught them, once a person (me) decided what to look for. The tool was never the problem. The missing question was.
I wrote four tickets, one for each fix. Each fix gets tested on a private copy of the app before members ever see it. Four problems, four written records. A meat proxy leaves no record of a decision, because no decision was made.
A meat proxy forwards output. A builder decides before generating, and owns what comes after.
You already have the hardest part. You know your customers better than any AI tool, and better than anyone on your team. That judgment is what’s scarce now, not the typing.
So here is what I would do this week. Don’t turn it into a test. The fastest way to get a bad answer is to put someone on their back foot.
I learned that years ago. When a team member told me there wasn’t enough time to do the job, I didn’t argue. We ran a time audit instead. For five working days, everyone, me and management included, used a time tracker to log what we worked on. It wasn’t exact, and nobody was judged. We just wanted to understand where the time went.
When we shared the results, with no names attached, we learned something.
Projected over 50 weeks, each person put in about 1,875 hours, and only about 768 were billable, roughly four out of every ten. Nobody would have offered up that number if I had asked them to defend their time. We found it because nobody had to.
Do the same with AI. For one week, everyone keeps a simple note for anything AI helped make: what it gave them, and what they changed. You keep one too. At the end of the week, sit down together and walk through a few. Use the four questions as a guide, not a grade, so you see where judgment is already happening and where it is quietly missing.
Then make one rule stick: decide what you want before anyone asks AI for it. The decision comes before the prompt.
If you pay someone outside to build or write for you, ask the same way, so you know what you’re paying for: “Help me understand how you use AI on my work.” A good vendor will be glad you asked. If they can’t tell you what they changed, you may be paying them for work the AI did.
And stop doing two things. Stop banning AI out of fear. And stop judging work by how fast it came back. Speed is the AI’s job. Judgment is the person’s.
The AI will always have an answer. Make sure someone on your side is asking the question.
Richard D. Stuart
The Weekly Prescription