In short In the Claude desktop app the same model works in three modes, because work comes in three "shapes." Chat — think turn by turn, message after message. Cowork — hand over a multi-step task whole and accept the result. Code — build and change software. Picking the shape that fits the task saves a lot of time.
Why three modes at all
Different tasks need different AI autonomy: sometimes you want a dialogue and control of every step; sometimes to hand off the work and get a finished result; sometimes to work on a codebase, where you need files, commands and edits. Doing everything in one mode is like hammering nails with a screwdriver.
Chat — turn by turn
The classic chat: you write — Claude answers — you refine. You hold the wheel at every step. When: questions, explanations, drafts, brainstorming, quick text — anything where the dialogue matters. Analogy: talking to an expert across the table.
Cowork — hand over the whole task
Claude works with your files, folders and apps: it plans a multi-step task and carries it out while you steer and accept the result. This is not a dialogue but a working session. When: "assemble a report from these ten files," "triage the mail and draft replies." Analogy: handing a task to a capable assistant. Full detail in Claude Cowork (needs a paid plan).
Code — build software
Claude Code is a coding agent that lives in the terminal and editors: it reads the codebase, runs commands, edits files. When: write a feature, find and fix a bug, refactor, review a pull request. Analogy: an engineer pairing right inside your project. Full detail in Claude Code — the terminal agent.
How to choose — cheat sheet
| Task | Shape |
|---|---|
| question, explanation, draft, idea | Chat |
| multi-step work with files, a finished result needed | Cowork |
| code: feature, bug, refactor, review | Code |
A sign you picked the wrong shape: in chat you're trying to assemble "over many turns" what should have gone to Cowork whole; or in Cowork you're splitting what was a simple question for Chat.
Related
The source is the Claude 101 course. Next — Claude Cowork and Claude Code — the terminal agent; how to frame a task in any shape — prompting and AI Fluency.