How AI Power Users Manage Context Libraries

A context library is a collection of knowledge that an AI agent can read or update. It may include markdown files, Notion databases, Slack history, or meeting transcripts. In a Nielsen Norman Group study, Claude power users spent more effort managing context than writing prompts. Most participants created their own system through experiments and advice from X or Reddit. Even these experts had missing files, old instructions, and unclear structures.

Context must change when the user’s work changes. However, a file may be written once and used for months. One participant had a daily task summary that slowly became less useful. Claude could access email, messages, and meeting notes, but the fixed instructions did not adapt. People also faced “system decay” when MCP or API connections expired. A broken Todoist connection forced one participant to reconnect it by hand.

Most participants did not edit context files directly. They asked Claude to make updates because it was faster and kept the style consistent. They also asked it to change many files when a project gained a new rule. But users rarely checked every edit. They judged the changes through later output. Because the agent created and maintained many files, the library’s structure often grew without a clear human plan.

People sometimes could not find important files in their own libraries. One participant found a global rules file that he did not know existed. This was acceptable when Claude could locate it, but poor structure creates risks. The agent may miss important context or load unnecessary information. Some users preferred local files in a terminal. Shared libraries were often backed up to Google Drive, while Claude’s cloud storage felt confusing to several participants.

The article suggests using an index file in each folder, often named CLAUDE.md, to explain its contents and use. The most organized library loaded as little as possible. The session’s starting folder limited what the agent could see. Context should also stay close to the project where it belongs. There is no perfect file length, but most files stayed under a few hundred lines. Let the agent help maintain the library, organize it with clear indexes, and load only what the task needs.