“Our AI expert wrote the prompt. Now we just package it up and the product is done.”
The one saying it is Tecla — who runs technology at the company, snorkelling mask on her forehead since this morning: the gear is always ready, the right moment never comes. The one answering is Rocco, the expert from the sun lounger next door, who already knows the next step. “Then it’s done. You put a button on top.”
The POC was real. A prompt that analyses the design of a machine: you give it the design and in ten minutes you have the analysis that used to cost half a day. No tricks, no prepared demo.
Then you open that prompt, to package it up, and you see what it actually is.
Inside there aren’t just questions. There are rules. If the material is this, check that. Never propose that solution below that threshold. Make sure the sum adds up. Written in plain language, in the middle of everything else, the way you’d write to a person.
And that’s the misunderstanding at the heart of the natural language revolution: because we can describe anything, we believe we’ve also imposed it. A rule written in plain language isn’t a constraint. It’s a well-worded request. As long as the right path is also the most probable one, it obeys. When it isn’t, it answers anyway — and it answers well.
Then someone adds the guardrails, which are more sentences in plain language. Which is to say: wishes, written more loudly.
In the POC it holds — and it only holds because of luck: there’s a person in front of the screen, one by one, quietly redoing whatever doesn’t add up without even mentioning it. A repetitive process running unsupervised has nobody in front of it. It runs at night, on project four hundred, and there’s no next shift to fix things.
Here comes the uncomfortable part for anyone selling AI: the missing piece isn’t AI. It’s basic software engineering. Thresholds, formats, the order of operations, what happens when a piece of data is missing, what must never come out — those are if-statements, tables, validations. Things we’ve known how to do for thirty years that nobody gets excited about today, because you can’t write them by talking.
The prompt stays, and stays where it’s unbeatable: the semantic part. Reading a specification, understanding what a drawing is asking for, pulling the meaning out of a document that no program could ever have read before. That’s the thing we couldn’t do before, and it’s enormous.
Everything else gets written in code, where a rule is a rule. Code that decides, prompts that understand: put together, they have a name — they’re called an agent.
You don’t package a prompt. You package a process, and inside it you put a prompt.