For the first couple of years, “prompting” meant one thing: writing a better sentence in a chat window. The lesson was to be specific, give examples, front-load context. That loop — write prompt, read output, revise — felt like the whole game. For a while, it was.
What wasn’t obvious at the time was that as AI moved from conversation partner to something closer to an autonomous worker, the old loop started showing its limits. Not because the models got worse, but because the conditions changed. When an agent runs for 45 minutes — or several hours — without checking in, the assumption that you’ll catch mistakes in real time quietly breaks. The skill you built for a back-and-forth conversation doesn’t transfer cleanly to directing work you won’t see until it’s done.
That is the shift worth naming. What becomes clear is that “prompting” now hides at least four distinct disciplines. Most of us are practicing one of them — probably the one we started with — and wondering why results feel uneven.
The four disciplines
The first is prompt craft — what most people still mean when they say “prompting.” Writing a clear, complete request; being specific; giving the model enough to work with in a single exchange. This matters. It’s also, in 2026, more table stakes than differentiator. Being very good at it gets you to par; it rarely gets you ahead.
The second is context engineering (the practice of curating and structuring all the information that surrounds your prompt in the model’s context window). A well-written prompt might be a couple hundred words, but the context window it lands in can be a million tokens wide — a million-token window is roughly the length of ten novels. Your prompt is a tiny fraction of what the model actually processes. Everything else — the files you’ve shared, the project documentation the agent can read, the memory of prior conversations — that’s context engineering. In practice, it means things like project-level instruction files, retrieval pipelines that surface the right documents, and agent memory architectures. The people who seem 10 times more effective with AI aren’t writing 10 times better prompts; they’ve built better context infrastructure.
The third is intent engineering, and this one is harder to teach because it’s less technical and more organizational. Context tells an agent what to know. Intent tells it what to want — the values and tradeoff hierarchies that should guide its decisions when specific instructions run out. There’s a documented case of an AI customer service agent that handled millions of conversations efficiently, cut resolution times dramatically, and then watched satisfaction scores fall. The agent was optimizing for speed because that was what it could measure. The actual organizational priority — relationship quality, when to bend a rule, when to spend an extra few minutes — was never encoded anywhere. A human employee with institutional experience would have known. The agent didn’t, because that knowledge was never given to it.
The fourth is specification engineering, and it’s the newest of the four. When an agent works autonomously over an extended stretch of time, the quality ceiling is set almost entirely by how well the work was described before it started. Not a prompt, not even a context document — a specification: what done looks like, how quality is measured, what constraints apply, where to escalate. The gap most people hit with AI — the output that’s almost right but not quite — is very often a specification problem rather than a model problem. A more capable model doesn’t fix an underspecified task; it just fails faster and more confidently.
Why the order matters
These disciplines build on each other. Good intent alignment requires good context. Good specifications require both. You can’t skip to the top and expect it to hold.
The genuinely interesting part is that this isn’t just an AI workflow observation. The discipline of writing complete, self-contained context — describing a problem as if the recipient has none of your background and no way to ask follow-up questions — turns out to make you a clearer communicator everywhere. Better at delegation, better at writing briefs, better at surfacing the assumptions that usually stay buried because shared context is assumed but doesn’t actually exist. One Shopify executive described it this way: what organizations call “alignment issues” is often just people talking past each other because the underlying assumptions were never made explicit. AI forces that discipline onto anyone who wants to use it well.
What’s next
The forward-looking question is: as agents run longer and more autonomously, which of these disciplines will matter most? My current answer is specification engineering, because the quality ceiling is set at the moment the work is defined — before execution begins. That’s a discipline most of us haven’t built yet. Most AI education still teaches prompt craft. Most practitioners are still perfecting that loop. The gap between those two facts is where the real leverage is right now.
Text summarized and optimized using Anthropic’s models and reviewed by a human.