Ask an AI chatbot to “write me a marketing email” and you’ll get something competent, generic and instantly forgettable. Ask the same model on the same day after giving it your brand voice guide, three examples of emails that performed well, a description of the customer segment and the objective of the campaign, and you’ll get something close enough to edit rather than rewrite. Same model, same task, completely different result — because the only thing that changed was what the model had to work with.

That gap now has a name: context engineering. Where prompt engineering was the celebrated skill of 2023 (finding the right words to coax a good answer out of a model), context engineering is the discipline of providing the right information: the documents, examples, constraints and history a model needs to produce excellent work, selected for the task rather than dumped in wholesale. It’s what separates people who get mediocre output from AI from those who get extraordinary output, and it’s now the most valuable AI skill a professional can build.

The shift feels familiar to me. I spent seventeen years running a search marketing agency, and early SEO went through its own version of this maturation. In the beginning it was a bag of tricks — keyword densities, meta tags, phrasing designed to catch an algorithm’s eye — and the tricks worked, for a while. Then the field grew up, and the practitioners who thrived were the ones who moved on from optimising sentences to building the right information environment: site architecture, content worth reading, authority earned over time. The tricks became table stakes and the discipline became the job. Prompt engineering is going through a similar transition, compressed into about three years.

Why prompts plateau

The reason is straightforward: a prompt can only rearrange what the model can already see. You can polish “write me a proposal” for an hour, and no amount of clever wording will conjure up your pricing structure, your past proposals, or the notes from last week’s client call, because the model doesn’t have them. We can hardly be blamed for starting with the wording (for a couple of years the internet was full of courses promising the perfect prompt formula), but once your request is clear, further work on phrasing returns almost nothing, while every relevant document, example and constraint you add returns a great deal. Andrej Karpathy, a founding member of OpenAI, endorsed the term in mid-2025 on the same grounds, describing the real work as “the delicate art and science of filling the context window” with exactly what the task in front of it needs.

What counts as context

Mostly it’s what you’d hand a capable new colleague on their first day. The source material that grounds the work in your facts rather than the model’s general knowledge: your style guide, your product sheet, the research report you’re actually discussing. Examples of what good looks like, because two or three strong ones teach a model your standard faster than any description of it. Standing instructions (what the trade calls a system prompt) that set role, tone and rules once, so you aren’t repeating them every session. The constraints that mark the edges: what it must never claim, the format the output has to arrive in, the things you’ve already decided. Structured data where it exists, because a tidy table of customer segments is far easier for a model to use reliably than the same facts scattered through prose. And, increasingly, access to your files, your data and your tools, letting it fetch what it needs rather than waiting for you to think of it.

Each of these has machinery behind it. Retrieval-augmented generation is the plumbing that finds and loads the relevant documents at the moment they’re needed; few-shot examples are the formal name for the show-don’t-tell move; memory management is deciding what carries over from one session to the next. You don’t need the vocabulary to use the ideas, though. Tools like NotebookLM turned document grounding into an entire product — upload your sources and every answer is anchored to them, something I’ve written about as a strategy copilot — while the project features in Claude and ChatGPT hold your standing instructions and curated files so that every new conversation starts already briefed.

Good output, in other words, is a briefing problem. Nobody expects a brilliant first draft from a new starter who’s been told nothing about the company, the client or the job — we’d call that a management failure. Yet that’s precisely how most of us use AI: a cold question typed into an empty box, then disappointment at the generic answer that comes back.

From job title to job requirement

The job market has already registered the shift. Companies that were advertising prompt engineer roles at $300,000 salaries in 2023 had largely stopped by 2025; Indeed’s search data shows interest in the role collapsing to a fraction of its peak, because the skill dissolved into everybody’s job description. What that looks like in practice is unglamorous: the account manager who keeps a folder of winning proposals to load before drafting a new one, the analyst who pastes the reporting template in alongside the raw figures, anyone who has learned to check the output against the sources they supplied before passing it on. That checking habit is the quality control the whole discipline rests on.

Where to start

Building the skill takes less time than you’d expect, and the habits compound quickly.

  1. Build a context library: collect the material you find yourself re-explaining every session — who you are, how you write, what you sell, examples of your best work. I’ve written a practical step-by-step guide to setting one up, and the whole job takes about ninety minutes.
  2. Package context by task: for each recurring job (the monthly report, the proposal, the campaign email), assemble a reusable bundle of instructions, source documents and a couple of strong examples. Load the bundle on its own; past a certain volume, extra context starts to degrade the output rather than sharpen it.
  3. Verify against the context: make checking the output against your supplied sources part of the routine — the numbers, the claims, the tone. Errors will still slip through from time to time; the difference is that you can trace them, because you know where every claim should have come from.

What I can’t tell you is whether context engineering will still be called that in three years. The tools are absorbing the discipline about as fast as we can name it — memory features, projects, agents that assemble their own working files — and it may go the way of spreadsheet literacy: a skill nobody lists on a CV because everybody is expected to have it. But the habit underneath the name will keep compounding, because the models will keep getting better whether we pay attention or not. The brief we hand them is the part that belongs to us — and learning to write a good one is worth more than any prompt formula ever was.