You use Claude Cowork by switching Claude Desktop into Cowork / Tasks mode, choosing a folder to share, and then describing the outcome you want in plain language. The official setup is straightforward: install and open Claude Desktop for macOS, select the mode selector (Chat vs Cowork), click Cowork to enter “Tasks,” and then tell Claude what to do. Cowork requires an active internet connection during the session, and the desktop app must remain open while the task runs. A good first task is something concrete and low-risk, like “Create a report.md that summarizes all Markdown files in this folder,” or “Organize these files into subfolders by type and month, and create a manifest.csv with old/new paths.” The important detail is that Cowork is designed for multi-step work: you give it a goal, it plans, executes, and produces finished artifacts directly in your file system.
To get reliable results, treat your prompt like a lightweight job spec rather than a casual request. Define scope (“Only use /Users/me/Projects/acme_notes”), constraints (“Do not delete anything; do not overwrite originals; write all outputs to /out”), and deliverables (“Generate summary.md, actions.log, and manifest.csv”). If the task changes files (renames, moves, edits), ask for a plan first: “Before making changes, list the exact file operations you’ll perform.” This mirrors how you’d run a safe batch script: preview → confirm → execute → log. Cowork can coordinate multiple workstreams in parallel for complex requests, but you still get better outcomes by breaking big jobs into stages (inventory, transformation, validation). For example, stage 1: “Scan and categorize files.” Stage 2: “Propose renames.” Stage 3: “Apply renames and write a log.”
Cowork becomes especially useful when you connect it to a downstream workflow rather than treating it as a one-off assistant. A practical pattern is: use Cowork to normalize and structure messy content, then ingest it into your systems. For example, have Cowork standardize documentation (consistent headings, front matter, deduplication), split long docs into smaller section files, and generate a machine-readable metadata.jsonl that maps each chunk to title/source/timestamp. Then your ingestion job can embed those chunks and index them in a vector database such as Milvus or Zilliz Cloud (managed Milvus). Cowork helps with the human-time-heavy prep work, while your pipeline stays deterministic: schema checks, embedding generation, and indexing remain in code where you can test and audit them.