Nielsen Norman Group: context libraries for AI power users
Context libraries for AI power users
Written by Tanner Kohler and published by Nielsen Norman Group on October 9, 2026, this article examines a practice growing among expert Claude users: building and maintaining “context libraries.” A context library is a collection of files — mostly markdown, sometimes connected to Notion or Slack — that an AI agent reads to understand how a person or team works.
The research found that participants spent more effort curating context than writing prompts. Despite that investment, many improvised their processes and let files go stale, which undercut the time savings agents were supposed to provide.
Three types of context
The article distinguishes three distinct roles for files in a library. Global context is stable information that shapes agent behavior across all tasks, such as communication preferences or writing style. Local context is scoped to a specific project or task, such as constraints for a particular design sprint. Ambient context is raw, uncurated input — emails, meeting transcripts — that the agent searches for relevant information rather than reading in full.
Blurring these types creates fragility: global instructions leak into task-specific work, and ambient data makes the whole library unwieldy.
Index files and findability
One unexpected finding: many participants couldn’t locate files in their own libraries. Some didn’t know certain files existed. The article recommends placing an index file — a CLAUDE.md “index card” — in each folder to describe its contents and tell the agent when to use it. This helps both people and agents navigate the library without loading everything into context at once.
Let the agent maintain the library
Rather than hand-editing files, participants described the change they wanted and asked Claude to update files — sometimes across many files at once. They judged results by output quality rather than reviewing each change individually.
External connections break often, which is worth building into maintenance expectations. Asking the agent to audit the context against recent work helps surface stale instructions before they cause problems.
Scope what each session can see
Starting a session in a specific folder limits what the agent can access, which prevents context from one project bleeding into unrelated work. One participant’s rule was that the agent must move up or down the folder hierarchy before moving sideways. Keeping individual files short — generally a few hundred lines — also helps the agent use them accurately.
Who it is useful for
Product designers and UX researchers who use AI agents daily, especially for writing, documentation, or systematic review work. The article is grounded in observed behavior rather than recommendations from first principles, which makes the tradeoffs visible. If you work with AI tools and have accumulated a collection of instruction files without much structure, this is a practical starting point for thinking about how to organize them.