If you’ve spent the last twelve months buying prompt-engineering courses, downloading 100-prompt PDFs, and copying frameworks off LinkedIn, and your AI output still sounds like it was written by a slightly tired temp, you’re not the problem. The prompt isn’t either.
So here’s the shift that’s actually happening, quietly, while half the internet is still selling prompt packs.
In 2026, Gartner identified context engineering as the breakout AI capability of the year, replacing prompt engineering as the discipline that separates the businesses getting real value out of AI from the ones still wrestling with mediocre output. Google Cloud has published guides on it. Atlan, Salesforce, and the enterprise context-platform tier have built their pitch around it. The Harvard Business Review put it cleanly in February 2026: when every company can use the same AI models, context becomes the competitive advantage.
That sounds like a lot of jargon. So let’s translate.
The simple version: prompts are the question, context is everything else
A prompt is what you type into the chat box. “Write me a LinkedIn post about my new offer.” That’s it. That’s the whole thing.
Context is everything the AI would need to know to answer that prompt well. Your voice. Your audience. What the new offer is, what’s in it, what’s not, what it costs. What you’ve launched before. What you’ve killed. The phrases you use, and the ones you’d rather walk into traffic than write. Your client stories. Your frameworks. The way you’d never call a customer a “consumer.”
Prompt engineering is the discipline of writing better questions. Context engineering is the discipline of giving the AI the right information, in the right format, at the right time, before the question even arrives.
One is the question. The other is the system that holds everything the question depends on.
Why prompts on their own hit a ceiling fast
The 54 percent of UK firms that say they use AI in some form (British Chambers of Commerce and Paragon Bank, September 2025) are mostly running on prompts. Around 11 percent have actually integrated AI in a way that produces consistent business value. The 43 percent in the middle is where most SMEs sit. Subscribed, dabbling, frustrated.
What’s happening in that middle? More often than not, it’s the same pattern.
You write a decent prompt. The output is technically correct, tonally wrong, and full of phrases you’d never use. You spend ten minutes rewriting it. You either give up on AI for that task, or you accept that you’re going to spend half your time editing. Either way, the productivity gain quietly disappears.
You can write a better prompt. You can write a longer prompt. You can stack three frameworks and a tone guide into a 600-word system message. It helps. A bit. Until you start a new chat tomorrow and have to do it all again, because none of that context persisted. The AI doesn’t know you. Every session starts with amnesia.
That’s the ceiling. And no amount of prompt engineering punches through it, because the problem isn’t the question. The problem is that the AI has no idea who’s asking it or for whom.
Recipe books, not Googling “cake”
Here’s the way I explain this on calls.
Using AI without a structured context layer is the difference between handing someone your grandmother’s recipe book and telling them to Google “how to make a cake.” Both will produce something cake-shaped. Only one will produce your cake. The one your customers actually buy.
Prompt engineering optimises the Google search. Context engineering hands over the recipe book.
If you’ve ever tried Custom Instructions in ChatGPT or Claude, you’ve already taken the first step. You wrote down a few things you wanted the AI to remember. A bio. A tone preference. Maybe a list of things to avoid. That’s useful. It’s also the difference between a Post-it note on the fridge and a franchise manual. The Post-it helps you remember the milk. It can’t run the business.
Context engineering, properly done, is the franchise manual. Structured. Governed. Versioned. Readable by every AI tool you use. Updated when things change. Pruned when they don’t.
Why this isn’t a passing trend
Context engineering is having its moment for a fairly unromantic reason. The maths has changed.
Every business now has access to the same frontier AI models. ChatGPT. Claude. Gemini. Copilot. The model itself stopped being a competitive advantage some time last year, because your nearest competitor can buy the same one for £15 a month. What you give it to work with has taken over the differentiator role.
McKinsey’s late-2025 AI report found that only 1 in 10 companies successfully scale their AI agents. The other 9 aren’t using worse models. They’re feeding them worse context. The pattern shows up at every size of business, from FTSE 100s down to the local accountant. Two-thirds of business leaders don’t fully trust their own data (Institute for Data and Enterprise AI). Two-thirds cite data silos as a top adoption challenge. Only 4 in 10 companies believe their data is AI-ready, and the number is going down, not up.
This is the structural problem context engineering solves. Not “make my prompts better.” Pay down the gap between what your business actually knows and what your AI can reach when you ask it something.
For SMEs, I call that gap context debt. The knowledge lives in the founder’s head, in a few decent Google Docs, in some old email threads, and in the shared muscle memory of two or three long-serving team members. AI tools can’t reach any of it. Every time you ask them to do anything substantial, they’re filling the gap with whatever they last picked up off the public internet. I ran the experiment recently, giving the same model the same brief twice, once with no context and once with a full context layer loaded. Same model. Same prompt. Two completely different businesses, apparently.
What context engineering looks like for a UK SME
Enterprise vendors talk about context engineering in language that doesn’t quite fit a five-person agency or a thirty-staff accountancy. So here’s the SME-sized version.
A working context layer for a UK service business of five to fifty staff has four jobs.
The first is to hold your identity. Your voice and tone, the way you actually talk, the phrases you’d never use, your brand assets, your personal profile as the face of the business. Without this, every piece of AI output is anonymous.
The second is to hold your business reality. What you sell, what’s in each offer, what you charge, who your ideal client is and isn’t, where you sit in the market, who your real competitors are. Without this, AI will quietly invent services you don’t offer and prices you don’t charge.
The third is to hold your evidence. The real client stories. The frameworks you’ve named and refined. The IP that makes you the choice over a generic competitor. Without this, AI will hallucinate case studies and mangle methodologies you’ve spent years building.
The fourth is to hold your operations. Your strategic priorities. What you’ve killed off. Your kill list. Your active launches. The templates you use. The regulatory context you work inside. Without this, AI will happily promote a service you shut down two months ago because nothing told it the news.
That’s not prompt engineering. None of it can be solved by a better question. It needs structure, governance, and a single source of truth your AI tools read from before they do anything.
The Firecore is context engineering for SMEs
What I build is called the Firecore. It’s an AI Context Layer, structured as four layers and fifteen modules, that sits inside the client’s own Notion workspace. They own it. I don’t host it. Every AI tool the business uses reads from it.
The Firecore is, simply, what context engineering looks like when you build it for a UK SME instead of a Fortune 500. The same discipline Glean and Writer sell to the enterprise tier at per-seat annual pricing that quickly runs into five figures. Built for a business with one founder, a handful of staff, and a desire to use AI properly without writing a six-figure cheque.
The Firecore is infrastructure. Not a chatbot, not a tool, not a prompt pack. It’s the layer that makes every tool, every prompt, and every AI session start from “the AI already knows your business” instead of meeting you for the first time, every time.
The honest answer to “can’t I just keep writing better prompts?”
You can. You’ll get marginal gains. They’ll fade the next time the conversation starts in a fresh window.
You could build the context layer yourself, too. WordPress is free. People still pay for a professional to do it properly. The architecture, the extraction process, and the ongoing governance are where the work actually sits. The technology is the easy bit. I went through this myself first, installing the system on Digital Magic Solutions before I’d sell it to anyone else. Four specific gaps surfaced. They’re the same four most founders find when they try.
The shift from prompt engineering to context engineering isn’t theoretical. Gartner has flagged it. HBR has named it the competitive advantage of the AI era. The enterprise tier has already moved. The SMEs that move next will out-compete the ones still buying prompt packs by the time the year is out.
Where to start
If you want a quick read on whether your business has the context debt I’m describing, the AI Context Layer Risk Quiz takes about six minutes. It walks through where your voice, your offers, your stories, and your strategic priorities actually live, and shows you the specific gaps that are costing you in AI output quality. Take the quiz here. You get a personalised score and a short report, no sales call required.
Either way, the question worth sitting with this week isn’t “how do I write a better prompt?” It’s “does my AI have anything to work with in the first place?” Get that right and the prompt almost stops mattering. Get it wrong and no amount of prompt engineering will move the needle.