Claude Desktop & Cowork — Tips & Tricks¶

Efficiency reference for Claude Desktop chat and Cowork mode. Source: Claude Help Center — Get started with Cowork, verified July 2026.

Projects vs. plain folders¶

Selecting a folder alone gives Claude file access but no persistent memory or instructions. To get memory, custom instructions, and scheduled tasks scoped together, create a Project.
Projects bundle: connected files/folders, custom instructions, memory files, and scheduled tasks into one persistent workspace.
Memory currently only persists within a project — standalone Cowork sessions don't retain memory across runs.

Global instructions (Settings > Cowork > Global instructions) — apply to every session, e.g. tone, default output format, your role.
Folder instructions — project-specific context tied to a connected folder; Claude can update these itself mid-session.
Project custom instructions — this project's persistent rules (what you're already using here).
Memory files — durable facts/feedback/preferences Claude writes and recalls automatically (see MEMORY.md pattern).

Claude writes discrete memory files (user, feedback, project, reference types) plus an index (MEMORY.md).
Give explicit correction/confirmation feedback ("don't do X", "yes, that approach was right") — this is what gets captured as reusable behavior, not just facts.
Use /remember style explicit asks for anything you want guaranteed saved.

A Skill is a repeatable recipe (e.g. a formatting standard, a report template) that keeps output consistent across runs instead of re-explaining format every time.
Trigger by name or let Claude auto-detect relevance from the request.
Combine multiple skills in one task when needed (e.g. docx skill + a styling skill).

Type /schedule in any Cowork task, or use the "Scheduled" sidebar tab, to turn a one-off task into a recurring job (daily briefing, weekly report, monitoring check).
Requires the Claude Desktop app open and the computer awake — scheduled tasks don't run if the app is closed or the machine sleeps.

Ask before acting — Claude pauses for approval on each action. Default for new tools/unfamiliar files.
Act without asking — faster, no pauses, but only use while actively supervising trusted files/sites.
File deletion always requires explicit "Allow" regardless of mode.

A plugin bundles skills + connectors + subagents for a role or workflow into one install. Faster than manually wiring up individual MCP connectors.
Check installed plugins/skills before assuming a capability is missing — ask Claude to list them rather than reinventing.

Dedicate a folder per project instead of pointing Cowork at broad folders like Desktop/Documents root — narrower scope = fewer permission prompts and safer file access.
Batch related work into one Cowork session; Cowork tasks consume more usage than plain chat, so avoid spinning up many small sessions for what could be one multi-step task.
Use plain Chat (not Cowork) for simple Q&A that doesn't need file access or long execution — cheaper and faster.
For inline edits on a drafted document: highlight the text, click "Edit with Claude," describe the change — avoids re-describing the whole document in the task thread.
Claude can coordinate parallel sub-agents on complex tasks automatically; you don't need to manually split the request unless you want explicit control over how it's divided.
Network egress and MCP connector permissions carry over from your Claude Desktop settings — web fetch/search and MCPs bypass egress rules, so review connector trust before granting access.

"Setting up Claude's workspace" on launch = normal version update, not an error.
Task stopped mid-run: check whether the Desktop app was closed or the machine slept — both end active Cowork sessions.
Windows "VM service not running": reinstall via the MSIX installer from claude.com/download, or sc start CoworkVMService.
Windows "EXDEV: cross-device link not permitted": set Settings > System > Storage > "Where new content is saved" back to C:\, then reinstall Cowork.