Chat, Cowork, and Code: which surface does your work live in?
One of the most common mistakes new Cowork users make is using the wrong tool for the job. They use Cowork to ask a quick question — and burn a week's worth of tokens on it. Or they try to use Chat to process a folder of 50 PDFs — and spend an hour doing manually what Cowork would do in minutes. Getting value from Claude starts with understanding which surface to reach for.
The three surfaces
Claude the intelligence runs on three different surfaces. All three are available on every paid plan. They are not separate products — they are different interfaces to the same Claude, optimised for different kinds of work.
| Surface | Interface | Who it's for | What it's for | |---|---|---|---| | Chat | Web and mobile | Everyone | Ask questions, think out loud, draft inline | | Cowork | Desktop app | Knowledge workers | Delegate outcomes, get finished files | | Code | Terminal (CLI) | Developers | Build software, run agentic development loops |
Chat: the conversation surface
Chat is what most Claude users start with. You type, Claude responds. You upload a file, Claude reads it and answers your question. You paste a draft, Claude edits it.
Chat is the right choice when the work is short, conversational, or doesn't need to produce a file. Quick research questions. Email drafts. Thinking out loud. One-off document edits. Chat is fast, light on token consumption, and available from any browser or phone.
The limit: Chat treats everything as a single conversation. You upload files one at a time. Outputs stay in the chat window — you copy and paste them into the actual document yourself. You are the orchestrator.
Cowork: the delegation surface
Cowork is Claude working on your files, not in a chat window. You grant it access to a folder. You describe an outcome. It plans the work, executes it autonomously, and delivers finished files.
Cowork is built on the same Claude Agent SDK as Claude Code — it is literally Code's agentic loop wrapped in a non-terminal UI. Anthropic built Cowork in approximately 10 days using Claude Code itself. That shared foundation matters: when Cowork handles a complex task, it brings the same planning, sub-agent coordination, and tool-calling depth that developers use in Claude Code.
The cost reality: because Cowork runs agentic loops — planning, tool calls, sub-agents, code execution — it consumes 5–10 times the tokens of a Chat conversation. This is not a bug. It's the cost of getting finished deliverables instead of text in a window. Use it accordingly.
Code: the developer surface
Claude Code runs in the terminal. It can read and write source files, run shell commands, execute tests, create commits, and build software end-to-end. If you're not a developer, this surface isn't for you — and you don't need it. Everything in this course uses Cowork.
The decision rule
When you're not sure which surface to use, apply this rule:
- Uploading one file and asking a question? Chat.
- Pointing at a folder and asking for a finished deliverable? Cowork.
- Building software or working in a codebase? Code.
The surface follows the work, not the other way around. A single file with a quick question is a Chat task even if the file is large. A folder with 20 files and a request for a synthesised report is a Cowork task even if the report ends up being short.
Key Takeaways
- 1Claude runs on three surfaces — Chat (web/mobile), Cowork (desktop app), and Code (terminal) — all available on every paid plan and all connected to the same Claude intelligence.
- 2Chat is for conversations, quick questions, inline drafts, and single-file work. You stay in the loop for every step.
- 3Cowork is for delegation: you grant folder access, describe an outcome, and Claude produces finished files autonomously.
- 4Cowork is built on the same agentic engine as Claude Code — it's the developer loop wrapped in a UI that doesn't require a terminal.
- 5Decision rule: one file, quick question → Chat. Folder access, finished deliverable → Cowork. Software development → Code. Cowork costs 5–10× more tokens than Chat, so match the tool to the task.