latest / Prompting is just writing

There’s a small industry now around “prompt engineering” as a discipline. Courses. Certifications. A vague suggestion that there’s a secret syntax, and if you learn the syntax the model will obey.

There isn’t a syntax. There’s just writing. The people who get good output from LLMs are, almost without exception, the people who could already describe what they wanted clearly in English to another human. The model doesn’t have a magic word it’s waiting for. It has a context window, and the question is whether you filled it with the things that matter.

The actually useful skills, in rough order:

• Naming the goal. “Refactor this to be cleaner” is a wish. “Refactor this so the public API has three methods instead of seven, preserving behaviour” is a brief.
• Naming the constraints. What can’t change. What’s load-bearing. What’s out of scope. The model can’t infer your invariants and will cheerfully violate them otherwise.
• Naming what good looks like. Show, when you can. Two lines of “here’s the style I want” outperforms two paragraphs of adjectives.
• Knowing when to start over. When a conversation has gone sideways, open a new one. The model isn’t getting smarter inside the bad context; you’re just compounding the misunderstanding.

None of this is exotic. It’s the same thing you do when you brief a contractor, or write a ticket that the person picking it up doesn’t have to ping you about. The thing the LLM is good at is reading carefully; the thing you have to be good at is writing things worth reading carefully.

The reason it feels like a new skill is that most of us have spent twenty years writing for compilers, which don’t care, and for Stack Overflow, which will correct us. The model neither cares nor corrects. It just does whatever the words said, slightly faster than you can think about whether the words said the right thing. Slow down. Write the brief. The rest is easy.