Claude: Context Management Remains a Challenge for Advanced Users

For daily users of Claude, the context library proves to be more challenging to conceive, store, and maintain than writing prompts. Between fragile connections, missing files, and updates delegated to the agent, certain practices emerge. They outline a continuous context engineering effort, often learned through trial and error, with concrete solutions to avoid obsolescence and overload.

Fragmented storage and uncertain tracking hinder usage

The cloud storage offered by Claude for context has been deemed unreadable and unreliable by participants. One user, after creating several agent skills in the desktop application, did not know where they were stored; another, faced with repeated failures, preferred to revert to Google Drive. Many favored working in the terminal, with easily accessible local files, and shared their library via Google Drive to allow for reading and modification by collaborators. GitHub was considered for version control, but no participant adopted it. During sessions, files deemed crucial were sometimes missing, to the point that some asked the agent to locate them. One participant discovered the existence of a significant global context file without prior knowledge. An incomplete understanding of the directory structure can lead the agent to miss important context not explicitly indicated or to load irrelevant information. The levels of organization varied greatly: some left Google Drive cluttered with disposable files created by Claude, while others maintained a meticulous hierarchy of over 2,700 files, or hundreds of folders nested within the Downloads folder. A method documented by AI labs involves placing an "index card" in each folder, describing its content and usage; in the Claude ecosystem, these files are named CLAUDE.md.

Continuous maintenance and brittle integrations

Context files were often written once and then used for months without being adapted to evolving practices. Some participants no longer checked the relevance of their context elements, while others continued to execute routines they no longer trusted or depended on fragile integrations with external databases. Unlike a human colleague, AI does not perceive implicit obsolescence, and enriching the ambient context, for example through MCP connections to meeting transcripts, did not remedy this limitation. Among the cases mentioned was that of a daily task synthesis automation that lost relevance as work methods evolved, even though the agent accessed emails, messages, and transcripts: the instructions remained too isolated and some information was omitted without explicit notification. A significant portion of the context came from external resources such as Notion or HubSpot, accessible via MCP or API, but the reliability of these accesses was problematic. Several participants expressed their frustration with managing these connections, with one even referring to a "system degradation": their Todoist link frequently disconnected, depriving Claude of access to the task list until a manual reconnection, to the point of considering an agent specifically tasked with maintaining connections. These obstacles call into question the expected time savings from using agents.

Let the agent edit, then judge based on the output

The majority of participants did not make modifications to their context files themselves. Even when the markdown file was open, they preferred to entrust the changes to Claude, believing this method was faster and ensured consistency in style and structure. Updates could be requested manually or triggered automatically by workflows, often involving multiple files at once, for example to integrate a new constraint into a project. Users did not always know which files needed to be modified or where they were located, and rarely checked the changes made by Claude. Instead, they evaluated the quality of the results produced by the agent, and in case of unsatisfactory output, requested a new round of targeted updates.

Keeping the library up to date: audit and explicit feedback

Concrete practices are emerging to limit context obsolescence: it is recommended to ask the AI to audit the global or local context in light of recent ambient context, for example by comparing the transcript of a meeting announcing a change in priorities with the project context, then validating and applying adjustments. Another key point is to provide explicit feedback to the AI when the way of working deviates from a routine: several participants knew how to articulate what was wrong but had not expressed it to the agent. Dictating this feedback and requesting the update takes a few minutes, a worthwhile investment for a frequently used tool.

What "context" encompasses and how users learn it

The context library encompasses institutional and procedural knowledge, often stored in markdown files, and can be extended by MCP or API connections to external sources like Notion, Slack, or Granola. Some information is actively updated, while others are only consulted. Three main roles are distinguished: global (stable and cross-cutting information), local (limited to a task or project), and ambient (uncurated raw flow). AI systems provide little guidance to users in managing context: many learn through posts on X, Reddit, or by experimentation, whereas mastering these tools should not require reading support pages intended for developers. Lessons from expert users working daily with Claude and possessing extensive libraries show that less technical audiences would likely encounter the same difficulties, amplified. In practice, participants devoted more effort to engineering these libraries than to writing prompts, often dictated on the fly, cobbling together personal methods; files became obsolete or disappeared, and many assumed that others were doing better.