Skip to content
Call wiki

Everything you need to know to start working with Claude Cowork

Last updated 2026-08-17 · Overview — no new claims; each statement links to the entry that carries the primary sources.

Cowork is Claude working as an agent on your files: point it at a folder and it reads, edits, and organizes documents over many steps, showing each one — no coding, any paid plan (Desktop vs browser). The mental model I teach: there’s an AI, and some text files you’re sending it to work on — those are the only two things that exist.

Where it runs, where files land

Section titled “Where it runs, where files land”

Cowork never touches your machine directly — it works in a cloud sandbox or a local virtual machine, and its outputs are only real files when you work in a folder you connected (Cowork sandbox). First move: pick a visible folder you can open yourself. The desktop app is what grants local-folder access; the browser matches it only for chat and cloud sessions (Desktop vs browser).

First-day setup

Section titled “First-day setup”
  • Connectors. Connect wherever your work lives — Drive, Slack, email, calendar. A connector sees everything the account sees, DMs included, and telling it to stay out of a folder is a request, not a boundary — keep must-never-touch data on an account Claude can’t reach (Connector scope).
  • Permissions. One ladder: approve each action, an auto mode where a second model reviews actions, or skip all checks; connector writes and org rules keep prompting on every rung (Permission modes). My workshop line is “lesson number one: always say yes” — what protects you is approval prompts, scope limits, and version history, not caution about trying things; prompt injection is heavily demonstrated in research but has no documented real-world personal-agent harm yet (Agent safeguards).
  • Dictation. Install Wispr Flow and talk instead of typing — people give far more context out loud (Wispr Flow).

Brief it like a newborn employee

Section titled “Brief it like a newborn employee”

Each session arrives brilliant but blank — “a newborn really smart baby, looking around trying to understand.” So over-share: dictate the purpose, the background, the format you want. Then make context findable: download the documents a project depends on into the folder first — live Drive search is a literal keyword query returning about ten results — and have Claude crawl once and write a navigation map (Drive search).

The mistake to avoid: trusting its reports

Section titled “The mistake to avoid: trusting its reports”

Cowork will lie to you — not maliciously, but a failed web fetch can silently become an answer fabricated from search snippets (Web fetch limits). Verify with concrete tests: demand a verbatim quote, fetch a known email, count the files. Have it work in stages and save every intermediate as markdown — agents edit Word and Excel blind, through code (Markdown vs Office) — then taste early: read the intermediates, not just the final output.

Habits that compound

Section titled “Habits that compound”
  • Keep durable state in instruction files and documents you can read back, not chat memory — the inclusion test is staleness, not importance (Memory vs files).
  • Delegate big: Cowork dispatches parallel subagents natively, each with its own context (Subagents in chat). Its time is cheap; yours is expensive.
  • Quality degrades gradually well before any limit — restart on quality drop or task switch, carrying a written summary (Context rot).
  • Run a low bar for starting things and a low bar for dropping them; when limits bite, upgrade the subscription before buying overage credits, which bill at full API rates (Plan economics).

When to move on

Section titled “When to move on”

Claude Code is the same engine on your real file system — a one-line install, driven in plain language; the choice is about wanting your own files and configuration, not coding skill (Cowork vs Code). The isolation you give up is a setting, not the product (Agent isolation).

Go deeper

Section titled “Go deeper”