In the grand schema of the new digital world of AI, data and systems; understanding matters, more than asking.
Everyone using AI right now has noticed the same thing: the results are inconsistent. Ask two people to use the same tool for the same task and you get two different qualities of output, and neither of them can tell you why. The usual explanation is skill, some people are better at prompting than others. That’s true, and almost beside the point. The deeper problem is that nothing compounds. Every session starts from zero. What one person figured out last week lives in their head, not in the system, and the next person has to figure it out again from scratch.
Context architecture is the name for treating this as a design problem instead of a skill gap, the same move code made for computers decades ago, applied now to what a model can see and act on. It runs across four cumulative levels:
Prompt craft is the single instruction: what you type in one box, one time. The Steering Wheel.
Context engineering is everything else the model can see while it works, the documents, the tool access, the prior exchanges held inside its context window.
Intent engineering encodes why the system is doing what it’s doing, not just what, so it optimises for the outcome you actually meant rather than the one you literally specified.
Specification engineering writes an organisation’s knowledge in a form an agent can read and act on directly, so a human isn’t translating intent into instruction every single time a task runs.
Each level depends on the one beneath it. A gap at any level caps everything built on top.
Go Deeper
- Next rung: What a Skill Is, where these levels stop living in your head and start living in files you own.
- The vocabulary: context architecture, context, datasets.
- Outside the vault: Anthropic on context engineering and Manus on the same lesson: the context window is the resource to manage.