Context Engineering Changes Everything About How We Work With AI

Something shifted in the AI world over the past year, and if you missed it, that's okay. Most people did. But it changes how all of us should be thinking about and using these tools.

The short version: it's not about writing clever prompts anymore. It's about giving AI the right information.

That shift has a name now. Context engineering. And honestly, once you understand it, a lot of the frustration people have with AI starts to make sense.

The Old Way Wasn't Working

For a while, the internet was full of "prompt engineering" advice. Write your prompt like this. Use these magic words. Say "act as an expert" and watch the AI get smarter. There were entire courses and certifications built around it.

And look, some of that helped. But here's what the research is actually showing us: the quality of your prompt matters way less than the quality of the information you give alongside it.

MIT's 2025 State of AI in Business report found that 95% of enterprise AI pilots delivered zero measurable ROI. Not low ROI. Zero. The root cause wasn't that the models were bad. It was that organizations weren't giving AI enough context to work with. They had powerful tools running on empty.

That stat stopped me in my tracks. Ninety-five percent. And the fix wasn't a better model or a fancier prompt. It was better context.

So What Is Context Engineering?

Andrej Karpathy, one of the most respected researchers in AI (co-founder of OpenAI, former head of AI at Tesla), put it in a way that clicked for me. He said in June 2025: "The LLM is a CPU, the context window is RAM, and you are the operating system." His original post on X kicked off an industry-wide conversation about what this shift actually means.

If that sounds technical, here's what it means in plain language. The AI model itself is the engine. The context window is its working memory, everything it can see and think about at once. And you? You're the one deciding what goes into that working memory.

That's the whole game. You're not trying to trick the AI with clever wording. You're trying to give it the right information so it can actually do good work.

By July 2025, Gartner made it official, declaring "context engineering is in, prompt engineering is out." When the biggest analyst firm in tech says the paradigm shifted, it's worth paying attention.

Why Bigger Isn't Always Better

Here's something that surprised me. Context windows (that working memory I mentioned) have expanded massively. We went from about 4,000 tokens a couple years ago to over 10 million tokens in 2026. You can now feed an AI an entire book and ask it questions.

But more isn't always better. Researchers found what they call the "lost in the middle" problem. When you dump massive amounts of information into a context window, the AI actually gets worse at finding and using the important parts. It's like handing someone a filing cabinet when they asked for a single document.

There's even research showing that long prompts degrade performance. The sweet spot? Roughly 150 to 300 words of well-chosen context. Not a wall of text. Not a one-liner. Just the right information, organized well.

This matters for everyone, not just developers. If you're using ChatGPT or Claude for your business, stuffing your prompt with everything you can think of is actually working against you.

What the Best Teams Are Doing Differently

Some of the most interesting lessons are coming from teams building AI-powered products.

Vercel (the company behind a lot of the web's infrastructure) shared a case study where they improved their AI's accuracy from 80% to 100% while using 40% fewer tokens. Not by upgrading the model. By improving what context they gave it. Less information, better chosen, dramatically better results.

Manus, an AI agent company, rebuilt their entire agent framework four times before getting context management right. Their big lesson? The way you structure and deliver context to an AI matters more than almost anything else. If you want to see how context flows between AI systems in practice, our explainer on MCP servers covers the protocol that's becoming the standard for connecting AI to external tools and data.

Anthropic, the company behind Claude, published a full guide in September 2025 on effective context engineering for AI agents. One of the key patterns they highlighted is what's called progressive disclosure: instead of dumping everything at once, you load information in tiers based on what's actually relevant right now. Give the AI what it needs for this step, not every step.

And here's a fun one that breaks a lot of the old prompt engineering advice: telling a modern reasoning model to "think step by step" actually hurts its performance now. Models like GPT-5 have internal routers that handle reasoning on their own. Adding that instruction is like telling a professional chef to "remember to turn on the stove." It just gets in the way.

The Human Touch Still Matters

Researchers at ETH Zurich studied something really specific: those context files that developers use to give AI background information about their projects (files like AGENTS.md or CLAUDE.md). What they found is that these context files genuinely improve AI performance, but only when they're written by humans.

When they had AI generate its own context files, the results were worse. There's something about human judgment, knowing what matters, what to emphasize, what to leave out, that AI can't replicate for itself yet.

I find that really encouraging. It means the skill that matters most isn't technical. It's the deeply human ability to understand what's important and communicate it clearly.

What This Means for You

Okay, so what do you actually do with all this? Here's what I'd suggest, whether you're a small business owner, a freelancer, or just someone trying to get better results from AI.

Stop chasing the perfect prompt. Instead, think about what information the AI needs to do a good job. Background, examples, constraints, tone, audience. That's your context. We covered this same principle from a different angle in why AI agents need better instructions — the quality of what you give an AI determines the quality of what you get back.

Be specific but concise. Remember that 150 to 300 word sweet spot. Don't write a novel, but don't be vague either. The goal is the right information, not all the information.

Give examples of what good looks like. If you want AI to write in your brand voice, show it three examples of writing you love. That's context engineering in action.

Layer your information. Start with the most important context. If the AI needs more, add it. Don't front-load everything. This is that progressive disclosure pattern the best teams are using.

Write your own context, don't let AI do it. If you're creating templates or instructions you'll reuse with AI, write them yourself. Your judgment about what matters is better than the AI's judgment about what matters. The research backs this up.

Update your context when things change. Old context is bad context. If your business, audience, or goals have shifted, update what you're telling AI. It can only work with what you give it.

We're All Figuring This Out

I want to be honest about something. Context engineering as a discipline is still new. The term barely existed 18 months ago. The research is coming in fast, and what we know today will probably evolve.

But the core insight feels durable: AI is only as good as the context you give it. That was true when these tools launched, and it'll be true for a long time.

The good news? You don't need a computer science degree to be great at this. You need clarity about what you want, thoughtfulness about what information matters, and the willingness to experiment.

That's it. That's context engineering. And honestly? It's way more accessible than prompt engineering ever was.

We're all learning this together. And that's kind of the whole point.

If you want help applying context engineering principles to your business workflows, our AI consulting team works with small businesses on exactly this kind of thing.

FAQ

What is context engineering and how is it different from prompt engineering?

Prompt engineering focuses on crafting the perfect wording to get a good AI response. Context engineering is broader — it's about curating the right background information, examples, constraints, and data that the AI needs to do its job well. Research shows that the quality of context you provide matters far more than the specific phrasing of your prompt.

How much context should I give an AI tool?

The research points to a sweet spot of roughly 150 to 300 words of well-chosen context for most tasks. Dumping massive amounts of information actually hurts performance due to the "lost in the middle" problem, where AI struggles to find the important parts in a sea of text. Focus on the right information, not all the information.

Do I need technical skills to be good at context engineering?

No. The most important skill is understanding what information matters for the task at hand and communicating it clearly. ETH Zurich research found that human-written context files outperformed AI-generated ones, which means your judgment about what to include and what to leave out is the skill that matters most.

Can context engineering help my business get better results from AI tools?

Absolutely. Vercel improved their AI accuracy from 80% to 100% while using 40% fewer tokens just by improving context quality. For a small business, this means better output from the AI tools you already pay for — better drafted emails, more accurate reports, more relevant content — without upgrading to more expensive plans or models.