Remember when “prompt engineering” was supposed to be the most important skill of the AI era? When people were sharing secret formulas and magic words to get better answers from ChatGPT?
That phase is mostly over. The models got smarter. The secret formulas matter less now. You can talk to them normally, and they will figure out most of what you mean.
But something else became important instead. Something that, honestly, should have been obvious from the start.
The skill that replaced prompting
A year ago, if you asked me what makes the biggest difference when working with AI, I would have said the prompt. The way you phrase your question, the structure, the role you assign. All of that mattered because the models were still rough around the edges. You had to work around their limitations.
Now? I can write a messy, half-formed question and still get a useful answer. The models improved that much. Prompting didn’t disappear. A clear question still beats a sloppy one. But the tricks stopped being the hard part.
What became the hard part is context. Not how you ask, but what the AI knows when you ask it. The best prompt in the world cannot save you when the context is wrong.
Why context matters more than you think
Here’s something easy to forget: when you talk to another person, context comes from everywhere. Your tone of voice, your facial expression, the shared history you have, the room you’re sitting in, even the time of day. All of it shapes how the other person understands what you’re saying.
An AI gets none of that. It gets text. That’s it.
If something isn’t written down in the conversation or in the files the AI can access, it simply does not exist for the model. It can’t guess your mood. It doesn’t know about the meeting you just had. It has no idea that when you say “the project,” you mean the one that’s been keeping you up at night, not the one you finished last week.
I think this is what most people miss. They treat AI like a colleague who shares their office, their history, their understanding of things. But the AI is more like someone who just walked into the room with no background at all. Everything you want it to know, you have to say.
The balance problem
Once people realize this, they often go to the other extreme. They dump everything into the AI. Full documents, long email threads, entire project specs. “Here, read all of this, then answer my question.”
And it works, sometimes. But there’s a catch. Even with today’s massive context windows (some models can handle over a million tokens), more is not always better.
There’s research showing that models pay more attention to the beginning and end of their context, and sometimes lose track of what’s in the middle. It’s called the “lost in the middle” problem. Pour in too much, and the important signal gets buried in noise.
Too little context and the AI guesses. Too much context and it gets confused. The skill is finding the right balance for each situation.
It’s project management, not engineering
Here’s where it gets interesting, at least for me personally.
I used to be a project manager. For years, my job was writing specifications, maintaining backlogs, making sure the team had everything they needed to do their work. I monitored progress, removed blockers, and kept documentation up to date. Not glamorous, but the project lived or died by how well that was done.
When I started working with AI agents more seriously, something clicked. The work I was doing felt familiar. Not the coding, not the prompting. The context building.
Every project I start now begins the same way. Before I ask the AI to do anything, I build the context. I write down what the project is about, what the goals are, what the constraints are, what has been decided so far. I set up the tools and connections the AI will need. Then, as the project moves forward, I keep all of it updated.
It’s not prompt engineering. It’s closer to being an architect or a project manager for AI agents. The same skills that made someone effective at managing a team of developers now make them effective at managing a team of AI tools.
Context is clay, not concrete
The biggest mistake I see people make is treating context as something you set up once and forget. Write the system prompt, add the instructions, done.
But context drifts. The project changes. Decisions get made in conversations that never make it back into the documentation. Three days in, the AI is still working from the original brief while reality has moved on. You get frustrated because “the AI keeps doing the wrong thing,” but the real problem is that you stopped updating its picture of the world.
I think of context as clay. You shape it at the start, but you keep reshaping it throughout the project. Add new decisions, remove outdated assumptions, adjust priorities. It’s maintenance work, and it’s not exciting, but it’s what keeps the AI useful over time instead of just in the first hour.
Why I need to see it too
One thing I feel strongly about: the context shouldn’t be a black box that only the AI reads. I need to see it too.
When your context is something readable, like a well-structured text file or a set of notes, you get two things. First, you can verify what the AI knows. If it gives a wrong answer, you open the context, look for the outdated part, fix it. No mystery.
Second, and this is more subtle: maintaining the context forces you to think clearly about the project itself. Writing things down in a way that both you and the AI can understand is a form of thinking. It’s the same reason writing documentation makes you find gaps in your own understanding.
This is why I prefer simple, transparent approaches over complicated automated systems. I would rather maintain a set of text files that I can read and edit than set up an automated pipeline that feeds the AI information I can’t see or control.
The hype around AI keeps moving. Last year it was prompting. This year it’s agents. Next year it will be something else.
But under all of it, the thing that actually determines whether you get good results is pretty basic: does the AI have the right information to work with? Is that information current? Can you see what it sees?
For me, the shift from prompt engineering to context management felt less like learning a new skill and more like recognizing an old one. Turns out, giving clear instructions, maintaining good documentation, and keeping your team informed works the same whether your team is made of people or machines.