Claude Power Users Spend More Effort on Context Than Prompts

A recent Nielsen Norman Group study found that experienced Claude users devote substantially more effort to engineering their context libraries than to writing the prompts that invoke them. Participants typically dictated prompts casually while improvising elaborate, often undocumented routines for maintaining the files, databases, and chat histories their agents draw from. The result, NN/g researchers report, is a fragmented setup in which context goes missing, grows stale, and duplicates itself across repositories.

What Is a Context Library?

NN/g defines a context library as a collection of institutional and procedural knowledge — stored in markdown files or external software — that an AI agent may reference and edit. The library functions as the agent's working memory: the corpus of project decisions, brand guidelines, coding conventions, meeting notes, and prior outputs that the model consults whenever it generates a new response.

Markdown files dominate the format, but the library extends well beyond them. NN/g researchers observed participants wiring up Model Context Protocol (MCP) servers and direct API connections to pull in external sources such as Notion databases, Slack histories, and Granola transcripts. Some entries are actively maintained; others are merely referenced when the agent needs them.

What Roles Does Context Play?

The analysis identifies three distinct roles that context plays within a library — a taxonomy NN/g treats in depth in a companion article. What the present piece makes clear is that the same chunk of context can shift between roles over time: a document that begins as a one-off reference may evolve into a standing rule once a team commits to enforcing it.

Why Do Even Experts Struggle?

The study surfaced a pattern practitioners will recognize: every participant had built a bespoke process. NN/g found that Claude power users rarely edit their context files once written and often cannot locate them when needed. Files proliferated across local folders, cloud drives, and chat threads, and many participants told the researchers they suspected someone else had solved the problem with a cleaner method — a perception that fueled a constant, low-level anxiety about their own setup.

The same dynamic applies to UX professionals, who treat context curation as a core design activity rather than a side task. The methodological implication is straightforward: if a single expert cannot reliably retrieve their own context, an agent that depends on that context cannot reliably retrieve it either.

What Does This Mean for Designers?

The findings frame context curation as a UX discipline, not a housekeeping chore. NN/g's analysis positions the context library alongside the information-architecture and design-system artifacts that already govern how teams store and reuse shared knowledge. Markdown's dominance signals that lightweight, text-based storage wins on portability and version control; MCP and API integrations extend that portability into the tools where institutional knowledge already lives.

The next iteration of agentic tooling will likely expose context libraries as first-class objects — searchable, auditable, and editable from within the chat surface itself, rather than as hidden inputs scattered across a user's filesystem.