AI Guides › Playbooks
By Nigel Guy · 6 min read
Most people pay for Claude and then use one door of it: the chat box. Everything gets typed in as a conversation, copied out, pasted somewhere else and tidied by hand. It feels fine because chat is good at a lot. It quietly fails when the job is really "go and do this in my files", "change this codebase" or "make me something I can look at".
The rule: pick the door by where the finished work has to end up, not by where you happen to be typing.
Claude now reaches you in four distinct ways. Names and layouts change often, so treat the labels below as the current ones, not permanent ones.
| Door | What it is for | The work ends up | Catch |
|---|---|---|---|
| Chat | Thinking, drafting, explaining, summarising, asking | In the conversation, for you to copy | You are the courier between Claude and the real thing |
| Cowork | Multi-step knowledge work: spreadsheets with working formulas, slide decks, formatted documents, recurring tasks | As files and finished deliverables | Uses more of your allowance than a plain question; local file and browser access needs the desktop app |
| Code | Reading a codebase, editing files, running commands, git | As changes in your project, with commits and pull requests if you ask | Built for developers; it can run commands, so permissions matter |
| Design | Visual work: prototypes, slides, one-pagers, designs | As a visual you can refine, then export | Labelled beta or research preview; some features are still basic |
Chat is the right door when the output is words and the next step is yours: an email draft, a second opinion, an explanation, a plan you will then act on. If you would be happy to read the answer and carry on yourself, stay here.
Cowork is Anthropic's term for bringing Claude Code's agentic approach to non-coding work. You describe an outcome and Claude works through the steps. Anthropic's help centre lists deliverables such as Excel spreadsheets with working formulas, PowerPoint presentations and formatted documents, plus cloud sessions that persist across devices and scheduled recurring tasks. It needs a paid plan (Pro, Max, Team or Enterprise). Local file access and browser automation need the Claude desktop app, and it needs an internet connection throughout. Multi-step tasks use more of your allowance than simple questions, and approval modes exist (manual, auto or skip) so you decide how much Claude may do unprompted. Anthropic's own guidance is that no mode replaces your judgement on sensitive work.
The tell that you should have used Cowork: you are on your fifth copy-paste from chat into a spreadsheet.
Code is Claude Code, an agentic coding tool that reads your codebase, edits files, runs commands and works with git. Anthropic documents it as available in the terminal, VS Code and JetBrains, the desktop app (the Code tab), and the web at claude.ai/code. A paid subscription or Anthropic Console account is needed for most surfaces. It is not only for professionals: if you maintain a website, a script or a small automation, this is the door that edits the actual files instead of describing the edit. Because it can run commands, read what it asks permission for.
Design is Claude Design, from Anthropic Labs: a way to create designs, interactive prototypes, slides and one-pagers through conversation, refine them with inline comments and direct edits, and export (Anthropic lists Canva, PDF, PPTX and HTML). Per the help centre it is in beta on Pro, Max, Team and Enterprise, switched on by default for Pro, Max and Team, and off by default on Enterprise until an admin enables it. You can reach it from a conversation, from the Artifacts tab, with /design in Claude Code, or standalone at claude.ai/design. Its usage counts against your normal plan limits. Known gaps at time of writing: inline comments sometimes do not persist, very large repositories can lag, multi-person editing is basic and there is no version history.
Run this in under a minute before you open a new task.
A routing prompt for when you are unsure:
You are a practical adviser on how to use Claude's different modes: Chat, Cowork, Code and Design.
My task: [TASK_DESCRIPTION]
What the finished result must be (file, code change, visual, or just an answer): [DELIVERABLE]
Where it must end up: [DESTINATION]
My plan and apps available: [PLAN_AND_APPS]
Anything sensitive in the files involved: [SENSITIVITY]
Do this:
1. Recommend one mode and give the reason in two sentences.
2. If the task needs more than one mode, give the order and what passes between them.
3. List what I should check or approve before Claude starts.
Rules: if any input is missing or I have not said which modes my plan includes, ask me before recommending. Do not state prices, limits or feature availability as fact; tell me to check Anthropic's current pages. Before answering, check that your recommendation matches the deliverable I named.
Fill in the five bracketed items.
Priya runs a small bakery and wants a wholesale offer. Her first instinct is to ask in Chat for "a wholesale pitch". She uses the table instead. Deliverable: a one-page price sheet plus a costing spreadsheet. The costing lives in a file she will update, so that is Cowork. The price sheet is something a café owner will look at, so that is Design. The email that goes with it is words she will send herself, so that is Chat. Three doors, each doing the job it suits, no copy-pasting of tables.