Study Maps Three Context Layers Power Users Build Around Claude

A recent Nielsen Norman Group study of Claude power users found that every participant had assembled a multi-layered "context library" of files, databases, and live data connections to steer the AI agent's output beyond generic responses.

The research identified three distinct roles of context that advanced operators curate: global context spanning multiple tasks, local context specific to one task, and ambient context drawn from raw streams such as email.

What are the three context roles?

In NN/g's taxonomy, global context carries information the agent should know across sessions or tasks — brand guidelines, organizational facts, persistent user preferences. Local context supplies the material for a single job: the brief, the source document, the target file. Ambient context streams into the agent from external sources such as a connected email inbox, supplying data the user would otherwise re-enter by hand.

The three categories differ in how they reach the model and how often they change. A static style guide behaves differently from a live inbox; treating them as interchangeable produces outputs that miss the brief or pull in irrelevant material.

How did the research surface this pattern?

NN/g worked with Claude power users — professionals who depend on the agent daily for substantial work — to map how they configured the tool. Every interviewee described maintaining a context library that Claude drew on alongside whatever the user typed in the prompt. Without those curated inputs, participants reported the agent produced generic or problematic outputs that did not reflect user needs, business goals, or a product's brand.

That finding inverts a familiar piece of advice. Curating the right context, the article argues, now matters more than writing a good prompt — a reversal of the prompt-engineering orthodoxy that defined the first wave of consumer generative-AI tools.

What does this change for product teams?

Teams shipping AI-agent features should treat the context layer as a product surface in its own right, with its own information architecture. Designers need to expose to users which global, local, and ambient sources an agent is currently holding, allow editing, and make it auditable which inputs contributed to a given answer. Long, opaque system prompts are harder to maintain than a structured context library with named entries and clear ownership.

For practitioners, the practical question is whether the current setup separates these three roles at all. Many ad-hoc uses of AI agents collapse global, local, and ambient context into a single prompt, which limits both accuracy and the ability to debug a bad output.

What does the taxonomy mean for the next wave of agent tools?

The three-role vocabulary offers a shared label for a design problem that has, until now, been solved ad hoc. As agentic features ship inside mainstream productivity tools, expect context curation to become a first-class workflow — carrying the same expectations for visibility, versioning, and governance that apply to any other production data source.