Context June 2026 8 min read

The end of prompts

For years we learned how to talk to AI: the right phrase, the clever trick. What truly matters has quietly shifted.

There was a time when prompt engineering was the hottest skill online. Courses, cheat sheets, whole collections of magic phrases. Find the right words and you got the better answer. Or so it seemed.

That time is ending. A simple insight has taken hold: how good an AI is depends mostly on what it knows about your work the moment you ask. And that is the part almost no one has under control.

The trick with the magic words

Early on, AI felt like a genie. Say the right words and you get what you want. So everyone collected phrases: act as a seasoned expert, think step by step, explain it like I'm a child. Some of it helped. Much of it was superstition.

For a short, self-contained question, that still holds. Summarise this text needs no craft. The moment AI takes part in real work, your project, your code, your decisions, even the finest phrasing quickly hits its limit.

The real bottleneck is the knowledge missing the moment you ask.

At any instant an AI sees only a slice of the world: whatever is right in front of it. Your language, it has. Your project, your decisions, where you left off yesterday, it knows only when they sit in front of it. What's missing doesn't exist for it.

The gap no one sees

That's where the uncanny feeling comes from that so many people know: the answer sounds right and yet isn't quite. The AI fills the gap with the most likely thing, because the specific thing is missing. And the likely thing often misses once it's about your case.

In the summer of 2025, Andrej Karpathy, a co-founder of OpenAI, gave this gap a name that has appeared everywhere since: context engineering.

The delicate art and science of filling the context window with just the right information for the next step.
Andrej Karpathy on context engineering, 2025

The difference is subtler than it sounds. Prompt engineering asks: how do I phrase the question? Context engineering asks: what must the AI know before the question is even asked? The developer Simon Willison, who defended the word prompt for years, pinned the problem down: too many people heard it as a fancy name for typing into a chatbot.

The myth of the bigger window

When the knowledge is missing, the first reflex is the obvious one: then just give the AI everything. The whole project, every document, the full history. Context windows have grown enormous, after all, millions of words fit inside.

Yet a bigger window doesn't bring better understanding. On this, the research is strikingly clear.

What gets lost in the middle

A widely cited Stanford study, Lost in the Middle, showed it back in 2023: feed a model many documents and it mostly uses the beginning and the end. Whatever sits in the middle drops away. Accuracy follows a U-shaped curve.

over 20%accuracy a leading model lost when the crucial piece of information slipped into the middle of the context.Liu et al., Lost in the Middle, Stanford 2023

In the worst case, the model did worse with many documents than with none at all. Researchers now call the effect context rot: the fuller the window, the less reliable the answer. Too much material can genuinely block the view.

A bigger context window gives the AI more room. Understanding doesn't follow from that.

Not the whole window, but the right slice is what counts.

Context still matters. What counts is the selection: the right thing, at the right moment, in the right form.

The real work

If selection is the real work, then context engineering is the name for it. And it's demanding. In practice, four moves have emerged.

Keep: save what matters so it stays available beyond the session.
Select: for each question, fetch exactly what counts.
Compress: summarise the long, strip the noise.
Separate: keep different tasks cleanly apart.

It sounds modest, yet at heart it's an architecture problem. It decides whether an AI agent stays reliable across many steps or loses the thread somewhere in between. The industry has noticed.

82%of IT and data leaders say prompting alone is no longer enough for serious AI.Industry survey, 2026

This very layer, between tool and model, is the work we put our time into at Thinkery. We're convinced this is where it's decided whether a powerful model becomes a reliable partner.

The question now is how well the AI is prepared for your question.

Where this leads

The end of prompts moves the work to where it belongs. With the right context handled in the background, no one has to wrestle with phrasing anymore. People can focus on judgement, on intent, on the decision.

This is where a tool becomes a real counterpart: the AI finally knows enough of your world to truly think along.

Good collaboration starts when both sides know enough about each other.

That's what we work on at Thinkery, day after day: a future where people and AI are a genuinely good team. Because the best answers come when the AI understands what you're really working on right now.