Prompt engineering is overrated. Context engineering isn't.

The prompt-tricks era taught a whole generation of professionals to write incantations. 'Act as a world-class marketing expert.' 'Take a deep breath and think step by step.' Templates with dozens of numbered instructions, traded around like rare cards. All of it rested on one assumption: that output quality lives in the wording of the question. That assumption was wrong, and it sent people hunting in the wrong place.

When AI gives you a mediocre answer, the cause is almost never a missing magic word. It's that you asked for specific work while giving the model nothing specific to work from. A perfectly phrased question about a document the model has never seen still produces a guess. People spent hours polishing the ask and zero minutes assembling what the ask was about, which is the same instinct that keeps teams stuck in the copy-paste commute: lots of motion around the chat box, no real material in it.

Here's the reframe. Brief AI the way you'd brief a capable contractor on day one. A good contractor doesn't need a secret phrase to do good work. They need to know what done looks like, who it's for, what to avoid, what good looks like, and they need the actual files. That's context engineering. Same brief, every time, for any model.

The field eventually said this out loud. In mid-2025, Andrej Karpathy endorsed 'context engineering' over prompt engineering, calling it 'the delicate art and science of filling the context window with just the right information for the next step,' and Shopify CEO Tobi Lütke defined it as 'the art of providing all the context for the task to be plausibly solvable by the LLM.' Simon Willison's explanation for why the new name stuck is the same diagnosis this piece makes: prompt engineering had collapsed in the popular mind into chatbot tricks, and the real skill got buried under them. In a 2026 survey of 250 IT and data leaders for DataHub's State of Context Management Report, 82% said prompt engineering alone is no longer sufficient.

Prompt tricks rot. Every time models change, some pet phrase quietly stops mattering, and the people who built their fluency on tricks have to relearn from scratch. The five-part brief doesn't rot, because it isn't a hack on any model's quirks. Goal, audience, constraints, example, sources: that's what any intelligent system needs to do unfamiliar work, silicon or human. Learn it once and it moves with you across every tool your company adopts next.

This stopped being a contrarian position sometime last year. Anthropic's engineering guide to context engineering treats a model's context as a finite 'attention budget' and defines the job as finding the smallest set of high-signal tokens that produces the outcome you want. Smallest is the operative word. Chroma's context-rot research tested 18 frontier models, including Claude Opus 4, GPT-4.1, and Gemini 2.5 Pro, and found that performance degrades as input grows, even on simple tasks, long before the context window is full. So the brief is a selection discipline: one gold-standard example instead of five, the three sources that matter instead of the whole folder, constraints that rule things out.

It also explains a pattern we see constantly at Candova AI: experienced managers often get fluent with AI faster than younger, more technical colleagues. Context engineering is delegation. Anyone who has handed a project to a new hire and watched it come back wrong has already learned, the hard way, that vague briefs produce vague work. They've been writing the five-part brief for years. They just called it managing.

The part of the brief where most people fail is the last one, sources. Describing your data to the model is not the same as giving it your data, and that gap deserves its own habit: upload first, ask second. Now the honest beat. Wording does still matter. Once the context is in place, a sharper instruction or a better-structured request will improve the result at the margins. We keep the useful parts of prompt engineering in the curriculum for exactly that reason. But it improves the last mile of a result whose foundation is already in place, and the individuals who train with us learn it in that order: context first, phrasing second.

The serious objection in 2026 isn't that briefing doesn't work. It's that the model increasingly does it for you. Agents now search your files, browse the web, and pull from connected tools on their own, and Anthropic expects agent design to keep moving toward 'letting intelligent models act intelligently, with progressively less human curation.' If the model can fetch its own context, why learn a framework for assembling it?

Because what's being automated is the fetching, not the deciding. An agent can find every document in your drive. It can't know which client this is for, what the legal team will reject, or which past report is the gold standard and which was a near miss. Goal, audience, constraints, example: four of the five parts of the brief are judgments only you can make, and they're exactly what delegating to a sharp human has always required. The fifth part, sources, really is getting easier, pointing instead of pasting, and that lightens your workload while leaving the skill intact.

There's a smaller objection worth a sentence: that context engineering is just prompt engineering with a new name. Fine. Call it whatever you like. The name changed because the old one had shrunk to magic-word tricks, and the work underneath, deciding what an AI needs to know before it starts, was always the bigger skill. The label will probably change again. The brief won't.