The best prompt I write is a folder. Here's how prompting got here.
In 2022, adding "Let's think step by step" to a prompt took one maths benchmark from 17.7% to 78.7%. Today OpenAI's own guide tells you not to bother. What replaced the trick is less clever and a lot more useful.
The trick era: When five words beat a bigger model
The first big lesson of prompting came with GPT-3 in 2020. You didn't need to retrain the model to teach it a task. You showed it three or four worked examples in the prompt and it copied the pattern.
Then, in 2022, researchers found something stranger. Google showed that examples with the reasoning written out roughly tripled a large model's score on school maths problems. A few months later a team in Tokyo found you didn't even need the examples. Adding "Let's think step by step" before the answer took an OpenAI model from 17.7% to 78.7% on one arithmetic test.
A fourfold jump from five words sent everyone looking for more of them. People opened prompts with "You are a world-class copywriter", offered the model tips, threatened it, and wrote rules in capital letters. Prompt libraries became a product category, and the advice spread faster than anyone tested it.
The hype: A job title with a $335k price tag
ChatGPT put all of this in front of everyone at the end of 2022. By March 2023 Fortune was reporting that Anthropic had advertised a "Prompt Engineer and Librarian" role paying up to $335,000. The number went everywhere, and for a while prompting looked like a career in its own right.
Anthropic still hires prompt engineers in 2026. The job has moved, though. The current roles are about writing the instructions behind agents and building the tests that check them. That's where the skill went.
The turn: The models learned the tricks
Reasoning models changed the picture. They work through a problem before they answer, so the step-by-step behaviour that once needed a magic phrase is now built in.
OpenAI's guidance for its reasoning models is blunt about it: "Avoid chain-of-thought prompts." The same page says to keep prompts simple and direct, use clear delimiters, and try without examples before adding any. Anthropic's Claude documentation opens on the same principle of clear, explicit instructions, and notes that newer models do what you ask quite literally. If you want them to go beyond the request, you have to say so.
So the most famous prompting trick of 2022 is now on a vendor's list of things not to do.
The rename: Prompting became context
In June 2025 Shopify's Tobi Lütke said he preferred a different term, "context engineering", which he described as giving the model everything it needs for the task to be plausibly solvable. Andrej Karpathy agreed a week later. His point was that people hear "prompt" and picture a one-line request, when real AI applications carefully build up everything in the model's context window, one step at a time.
Anthropic's engineering team followed with a long piece on the subject in September 2025. They called context engineering the natural progression of prompt engineering, and they described context as "a finite resource with diminishing marginal returns". The more you load into a model, the worse it gets at recalling any particular part of it. The aim is the smallest set of useful information that gets the job done, not the most.
You'll find plenty of posts this year declaring prompt engineering dead. I don't think that's right. The wording still matters. It's just become one part of a bigger job, which is deciding what the model knows when it starts work.
My version: The prompt is a folder
I wrote earlier this year about running the business from a folder of markdown files. What I didn't appreciate when I built it is that the whole thing is one long prompt, split into pieces that load only when they're needed.
The root file is about 1,500 words. It doesn't try to tell Claude everything. Most of it is a table that says when to go and read something else: the brand file for design work, a client's file for that client's work, the invoice procedure before touching an invoice. Anthropic's article calls this just-in-time retrieval. I got there for a duller reason: anything that loads every time is something I have to keep correct every time.
Below that sit 19 skills, each a markdown file describing one workflow, and 99 short memory notes Claude has written over the months. Most of those notes carry two lines that matter more than the fact itself: why it's true, and how to apply it. The reason is what lets the model handle the case I didn't predict.
None of these files contain a persona or a magic phrase. They read like instructions you'd leave for a capable new starter.
The lesson: Explain the why, then get out of the way
The biggest change in how I write instructions is that I explain them. My blog skill asks for meta descriptions of 140 to 150 characters, and it says to measure them with a short script rather than by eye, because models miscount characters. Without that reason, the script looks like fuss. With it, the model knows its own counting is the weak spot and stops trusting it.
I still use examples, but to show the output I want rather than to teach reasoning. One good example of a finished headline does more than a paragraph describing one.
The other habit is pruning. A rule that no longer applies still costs something. It sits in the context, takes attention away from the rules that matter, and occasionally gets followed when it shouldn't. Going through the folder and deleting things is part of the job.
For your business: Write it down once
You don't need a folder of markdown files to use any of this. ChatGPT and Claude both let you set up projects with standing instructions and reference files, and they're easy to leave empty.
Start with what you'd tell a new member of staff on their first morning: who your customers are, how you talk to them, what you sell and what you never promise. Put it in once and reuse it. When you ask for something, say plainly what you want, why you want it, and show one example if format matters. Then check what comes back, because the model will still get things wrong.
The bottom line: Magic words are out, good briefs are in
Prompting started as a hunt for magic words and has turned into something closer to writing a good brief. The models caught up with the tricks, and what's left is knowing your business well enough to explain it.
No model can do that part for you.
Steve Lavine is a full-stack developer and founder of Lavine Web & AI Solutions, working with SMEs and start-ups across the UK. lavine.dev