Detailed Analysis
A Reddit user's account of Anthropic's merger of "Cowork" (Claude's desktop application feature that grants direct filesystem access to local project vaults) with the standard web-based "Projects" feature has surfaced significant friction in the transition. According to the report, Anthropic updated the desktop app interface to unify these two previously distinct workflows, but the migration produced unintended duplication: existing projects were cloned into parallel "Cowork" versions, each carrying over the original project's files and custom instructions. Because Cowork's entire value proposition is direct, live access to a user's local vault, this duplication created redundant and conflicting file sets—project instructions written for the web-based workflow now collide with the CLAUDE.md configuration files that govern behavior within the Cowork vault itself.
The practical impact described here is instructive: rather than a clean feature upgrade, users are confronted with what amounts to a data-integrity and configuration-management problem. The original poster reports losing roughly a third of a work session simply diagnosing the change before arriving at a workaround—writing project-level instructions that explicitly tell Claude to detect whether it's operating in Cowork mode with direct vault access and, if so, defer to CLAUDE.md rather than the (now-redundant and possibly contradictory) web project instructions. This is a nontrivial burden to place on end users: it requires them to author conditional logic in natural language just to keep two overlapping context systems from stepping on each other, and it surfaces a deeper issue about how Anthropic communicates and stages changes to power-user tooling that people have built real workflows around.
This incident matters beyond one user's inconvenience because it touches on a recurring tension in AI product development: the push to unify and simplify interfaces often collides with the reality that sophisticated users have already built intricate, interdependent systems atop the existing structure. Cowork's local-vault access model represents a meaningfully different paradigm from cloud-hosted Projects—one is essentially agentic filesystem access, the other is a curated document/context store—and collapsing them into a single interface without a carefully engineered migration path risks exactly the kind of instruction-conflict and duplication described. For a tool that many users rely on for real work (note-taking systems like Obsidian-style vaults, coding projects, research repositories), silent restructuring of file organization is a serious trust issue, not merely a cosmetic UI change.
More broadly, this episode reflects the growing pains of AI companies rapidly iterating on agentic and context-management features—memory systems, project scaffolding, file access, and persistent instructions—without fully mature versioning, migration, or rollback tooling. As Claude and competing assistants move from stateless chat interfaces toward persistent, file-aware, agentic systems that maintain long-running project state, the underlying infrastructure for managing configuration conflicts, deduplication, and safe schema migrations becomes as important as the AI capabilities themselves. Complaints like this one function as informal user-driven QA, signaling to Anthropic (and to the broader industry) that feature convergence needs to be paired with robust conflict resolution and clear user communication—otherwise the cost of "innovation" is quietly offloaded onto users who must reverse-engineer what changed and patch it themselves with ad hoc prompt engineering.
Read original article →