Building AI for your own work has a blind spot that's easy to miss. You build the system, you look over the output, it looks right to you. But your work rarely stops with you. It goes out to someone, and they might see something you can't.

​

There's a habit worth identifying. “Prompt-and-paste.” Someone takes an AI output and drops it straight into the work. It can be a whole deck, a section of a document, an email, whatever. Prompt in, copy the output, paste it. Sure, they’ve read it and it seems fine, so they send it on. What they don't see is the person on the other end, who can tell. We've all seen AI outputs for a few years now, and most people can feel this sort of output the way they can feel a stock photo.

​

So what exactly is it the people on the other end are feeling? In part it’s the language. But, really, it's the flatness of the output. There’s a lot of words and no point. No perspective. The reader gets nothing out of it because there wasn’t anyone in the system designing what actually mattered. Left to its own, AI weighs everything the same. A person doesn’t do that, whether they intend to or not. They go deep on some things and lighter on others. The reader might not agree with what they’re reading, but at least there’s something to agree or disagree with.

​

The fix isn't a better model or an article about prompting. It's building steps into the system that makes the AI challenge its own outputs. Even with great knowledge documents and system prompting, AI is going to give you obvious answers. And if you say, “come up with unconventional thinking” you risk outputs that aren’t grounded in reality. So you build in one pass that does the logical thinking and another pass that challenges it to find where it’s generic and flat. Or you go the other way and have it create boundless, “what if” thinking first, and the second pass to ground things in reality. That's how a point of view gets built. People do it all the time and it’s what happens when people work. Someone throws out an idea, another person challenges or riffs or takes it somewhere unexpected. And a third person finds the thread that makes it all work. AI systems need to work the same way.

​

Just like a human brainstorm session, an AI system can only challenge itself if it knows what it’s aiming for. What aspect of an idea is it pushing back on? What does generic mean? The system needs standards, and that’s defined by the team. What are the parts that matter, what parts can be ignored and what does ‘good’ look like. It takes time to dissect this and structure a system to fold it in. But when you have those and they’re built into the system, it works from the same judgement you do.

​

Build with AI. Build it around your work and your thinking. Build it around the thinking behind your thinking. But don’t hand over your work to it without doing this first. The second you do, the work goes flat, and the person on the other end can feel it before they've finished the first line.