The Sentence I Now Say in Every First AI Session

Last updated: June 2026

Priya had done her homework before our first session, which is usually a good sign and this time was the problem. She was a product manager at a logistics company, mid-thirties, sharp, and she had spent two weekends working through a prompt engineering course. She showed up with a notebook full of templates. Role assignments, delimiters, few-shot examples, the whole vocabulary. She could recite the difference between zero-shot and chain-of-thought.

And her Claude outputs were still mediocre. That was why she had booked the call.

She pasted a prompt she had labored over, a careful thing with a "You are a senior supply chain analyst" preamble and three formatting instructions. The answer came back generic. She sighed and said the thing I have now heard some version of from dozens of professionals. She said she must be doing the templates wrong.

I asked her to do something that felt backwards to her. I asked her to forget the templates and just tell me, in plain English, like she was talking to a coworker, what she actually wanted.

The moment it clicked

She talked for ninety seconds. No structure, no role-play, no magic words. Just: I have a spreadsheet of late shipments, I need to find the pattern, I think it's regional but I'm not sure, and I have to explain it to a VP who hates jargon.

I told her to paste exactly that into Claude, then add one line: before you answer, ask me five to ten questions that would help you give me a better answer.

Claude asked her eight questions. Two of them were about things she had never thought to include, like whether weekends counted as delays in her data and which carriers she used. She answered them. The next response was the analysis she had been trying to engineer for two weekends, and it was better than anything her templates had produced.

She went quiet for a second, then said the line I now think about constantly.

"So I don't have to get good at prompting. I have to get good at telling it what I actually want."

That is the whole thing. She had spent her weekends learning to speak the model's language, and the model had quietly gotten good enough that it would rather just speak hers.

Why this is the pattern, not the exception

Priya was the third professional that quarter who came to me frustrated by a prompt engineering course. The pattern is always the same. Smart person, real intent, buried under a layer of memorized technique that the current models no longer need. The clever-phrasing era was real, around 2023, when you genuinely had to coax the model with the right scaffolding. That era is closing. Anthropic's own tooling now ships a prompt generator that writes the engineered prompt for you from a plain description of your task, which tells you where the company itself thinks the skill is heading.

What replaces prompt engineering is not nothing. It is the opposite skill. Instead of compressing your intent into the model's preferred format, you expand your intent in your own words and let the model interview you for the rest. I have started calling it meta prompting, because the move is to prompt the model about the prompt: ask it to improve your draft, ask it to ask you questions, ask it what context it is missing.

The professionals who pick this up fastest are not the most technical ones, which surprises people. Priya was good at it because product managers spend their careers turning vague stakeholder wishes into clear specs. That muscle transfers directly. The same instinct shows up when I teach people to work with Claude at their job or when an analyst first reaches for AI instead of a colleague who knows ChatGPT and coding tools. The bottleneck was never the syntax. It was always the willingness to say plainly what you want, and to admit out loud what you are unsure about, so the model can fill the gap.

What I tell every student now

I open almost every first AI session the same way these days. I tell the student we are not going to memorize a single template. I tell them the model already knows more about prompt structure than any course will teach them, and that their job is the one thing the model cannot do for them, which is to know what they actually need. Then we practice two moves until they are automatic: paste your rough draft and ask the model to make it better, and ask the model to ask you questions before it answers.

It feels too simple to people who have already paid for a prompt engineering course, and I understand the resistance. Letting go of the templates feels like giving up a skill you just bought. But the result is consistent enough that I no longer hedge about it. Priya emailed me three weeks later with a dashboard analysis she had run entirely on her own, and the only "technique" in it was a plain paragraph of intent followed by a request for clarifying questions.

If you have a notebook full of prompt templates and outputs that still disappoint you, this is the broader argument I make for working with AI as a professional. It is also the thing a student of mine asked me for directly when he said he wanted me to teach him to think in the AI way rather than hand him another list of tricks. If you recognize yourself in Priya's story, a free Discovery Call is where hers started.