Detailed Analysis
A Reddit user's report of Claude Projects and Skills failing to sync between office and home devices highlights a friction point that sits at the intersection of Anthropic's rapid feature rollout and the practical realities of hybrid work. The user describes building a custom skill that allows Claude to access a Dropbox folder of PDFs, read through them, and execute tasks like drafting reports, renaming files, and writing letters. After editing that skill during an office session and explicitly instructing Claude to "save this change to the skill," the user returned home to find neither the project nor the updated skill was accessible — the skill showed a timestamp over a week old, and the project itself wasn't visible at all.
The likely technical explanation, which the user themselves speculates about, points to a common failure mode in how conversational AI systems interpret persistence instructions. When a user says "save this change to the skill," there's ambiguity between updating a session-scoped or project-scoped state versus writing back to the actual underlying skill file that should propagate across all instances and devices. If Claude interpreted the instruction as the former, the edits would appear to "stick" within that single working session but never actually modify the canonical skill definition that syncs elsewhere. This is a classic problem in stateful AI tooling: the model may confirm an action verbally ("I've saved this change") without the underlying infrastructure actually persisting or replicating that change to a shared, cross-device store.
This matters because Anthropic has been pushing Claude increasingly toward being a persistent, agentic work tool rather than a stateless chatbot — Projects and Skills (along with Artifacts, memory features, and Claude Code) are all part of a broader strategy to make Claude behave more like a continuous collaborator that retains context, tools, and custom configurations over time. For that vision to hold up under real-world hybrid or multi-device workflows, state management needs to be rock solid and behave predictably: what you save on one device should reliably appear on another, and the model's natural-language confirmations need to accurately reflect what actually happened in the backend. A gap between what Claude tells the user ("saved") and what the system actually persists undermines trust in exactly the agentic capabilities Anthropic is trying to showcase — file manipulation, task execution, and long-running project work.
More broadly, this incident reflects a recurring theme across the AI assistant industry: as products layer increasingly complex memory, tool-use, and customization features on top of large language models, the seams between "the model says it did X" and "the infrastructure actually did X" become more visible and more consequential. Sync issues, ambiguous persistence semantics, and inconsistent state across devices are exactly the kinds of unglamorous but critical engineering problems that will determine whether AI assistants can be trusted as genuine productivity tools for professionals who move between multiple machines and locations — a use case that is now the default rather than the exception for much of the knowledge-work population.
Read original article →