Host: Have you noticed how many courses there are lately on how to 'properly' talk to AI? It's like we've created this whole new career path called prompt engineering just to get software to do what we want.

Listener: Oh, absolutely. My feed is full of 'prompt libraries' and templates. I actually felt a bit guilty for not studying them. I figured I was just holding the tool wrong.

Host: That's exactly what the industry wants you to think. But there's this provocative argument out right now that prompting isn't actually a skill gap in users—it's a symptom of unfinished products. Basically, the software is making you do its homework.

Listener: Wait, so you're saying if I have to learn a specific 'grammar' to get an answer, the software is the one failing, not me?

Host: Precisely. Think about any other software. If it doesn't work when you click a button, that's a defect. But with AI, we've turned that gap into a curriculum. And here's the kicker: even the experts who think they're good at this might be wrong about how much it's helping.

Listener: What do you mean 'wrong'? If I'm using a prompt and I feel faster, isn't that proof enough?

Host: You'd think so, but there was this study by METR in 2025. They took sixteen experienced developers—people who knew their code inside and out. These experts predicted AI would make them twenty-four percent faster. After the work, they felt about twenty percent faster. But when the researchers actually checked the stopwatch? They were nineteen percent slower.

Listener: Nineteen percent slower? That's a huge gap between feeling productive and actually being productive. Why the disconnect?

Host: It's what the report calls the 'thirty-nine point gap.' When a skill is that hard to measure, it's indistinguishable from just a habit you've grown comfortable with. The product becomes 'unfalsifiable' because you can always blame the user's prompt instead of the tool's output.

Listener: That's a bit unsettling. But if we aren't supposed to prompt, how are we supposed to get the different AIs to do what they're best at? I mean, I've heard Claude is better for some things and ChatGPT for others.

Host: That's another point the research makes. In 2025, ChatGPT traffic was over seventy percent non-work related, while over a third of Claude conversations were technical or mathematical. But why should you have to research which 'leaderboard' is winning this week? Matching a task to the right model is a technical problem that software should absorb, not something you should have to figure out.

Listener: Okay, I get the theory. But in the real world, I'm usually rushing between meetings. I don't have time for the software to 'absorb' anything if I can't just tell it what I need.

Host: Actually, your stress is the strongest argument for this. Microsoft did a study showing the average knowledge worker is interrupted every two minutes. And get this: editing on presentation decks spikes one hundred and twenty-two percent in the final ten minutes before a meeting.

Listener: Okay, that's definitely me. Nine minutes to go, three chat windows open, trying to fix a slide.

Host: Right! In that moment, you aren't going to craft a 'well-structured prompt with clear role definitions.' You're going to type what you want in whatever words you have left in your brain. If the software needs you to be a 'calm, well-briefed user' to work, it’s designing for a person who doesn't exist during work hours.

Listener: So what's the alternative? If I just shout a messy sentence at the computer, how does it actually get the work done right?

Host: The report suggests the system itself should do the heavy lifting. If a request needs to be broken into steps, the system should decompose it. If it needs a plan, the system should plan. That's the idea behind things like Aura and Echo OS—you state your intent in plain words, and the software handles the sequencing and the checking under the hood.

Listener: It sounds like we're moving from 'you have to learn the machine's language' back to 'the machine finally learns ours.'

Host: Exactly. The real measure of an AI product shouldn't be how much you can extract from it once you've mastered its complex grammar. It should be how little you had to learn to get it to work. If you want to see the full breakdown of that productivity study or the workplace stats, the full report is definitely worth a read.