// The guide
Which Claude, when? Every surface, one decision guide.
Claude used to be a chat window. In 2026 it's a chat, an agent workspace (Cowork), a coding agent (Claude Code — CLI, desktop and web), add-ins inside Excel, Word and PowerPoint, a browser extension, and a teammate in Slack. Same brain, eight bodies.
That's power — and confusion. Pick the wrong surface and you'll burn an afternoon copy-pasting between windows, or burn your usage quota doing in chat what an agent does in one command. This guide is the map.
The one mental model that sorts everything
Every Claude surface answers one question: where does the work live?
- Work lives in a conversation → claude.ai chat
- Work lives in your files and apps, and you want it DONE → Cowork
- Work lives in a folder you version and automate → Claude Code
- Work lives in one specific document → the Excel / Word / PowerPoint add-ins
- Work lives on websites → Claude in Chrome
- Work lives in your team's channels → Claude in Slack
Same engine underneath. The surface decides what Claude can touch.
The decision table
| You want to… | Use | Why not the others |
|---|---|---|
| Ask, brainstorm, draft quick text | Chat | Agents are overkill; chat is fastest and cheapest on quota |
| "Take these 30 receipts and build the expense sheet" | Cowork | Chat can't touch files; Code is dev-flavored for the same engine |
| Organize a folder, run a weekly ritual, automate | Claude Code | Versionable, scriptable, schedulable — see the non-dev guide |
| Fix a formula, debug #REF!, change a model assumption | Claude for Excel | It cites cells and preserves formula dependencies; nothing else does |
| Review a contract in tracked changes | Claude for Word | Native revisions, comment threads, redline synthesis |
| Slides that respect the corporate template | Claude for PowerPoint | Template-aware, native editable charts |
| Fill web forms, extract from tabs, automate a web task | Claude in Chrome | Only surface that clicks and types in the browser (beta) |
| Delegate inside team discussions | Claude in Slack | @Claude, visible to the whole channel (Team/Enterprise) |
What most people get wrong
1. Doing agent work in chat. If your prompt contains "then… then… then…", you're describing a Cowork session or a Claude Code task. Chat will give you the pieces; you'll assemble them by hand. The agents assemble.
2. Ignoring that the quota is ONE pool. Chat, Claude Code and Cowork share your 5-hour and weekly usage limits. One sprawling Cowork session costs as much as dozens of chat messages. Corollary: don't burn agent quota on questions, don't burn chat turns on multi-step work.
3. Long dirty sessions. Every turn re-reads the whole conversation. A 200-turn chat where you changed topics four times costs multiples of four clean sessions. New topic, new session — your quota will thank you.
4. Using computer-use where a connector exists. The hierarchy that Anthropic itself recommends: connector first, dedicated tool second, screen control last. Fast and precise beats general and slow.
5. Treating the surfaces as separate products. They compose: research in chat → build in Claude Code → verify in Chrome → present via PowerPoint. The conversation even follows you across Excel, Word and PowerPoint — read the numbers in one, write the slides in the other, no copy-paste.
Plans, in one paragraph
Free gets chat with web search and memory. Pro (from $17/month annual) unlocks the real toolbox: Claude Code, Cowork, Research mode, the Microsoft 365 add-ins, unlimited Projects. Max buys 5-20x the capacity — it's the plan for people who live in the agents. Team/Enterprise add the admin layer, and Slack's Claude Tag lives there. The differences are mostly capacity, not features.
The thread that ties it together: memory
Eight surfaces, one recurring frustration: each session starts from zero about YOU. claude.ai's memory (open to everyone since March 2026) remembers chat preferences; Claude Code keeps notes per project per machine. But your decisions, your clients, your way of working — the stuff that makes an assistant a right hand — still evaporates between sessions and surfaces.
That's the layer we built Klyr for: a persistent memory of who you are that lives inside Claude Code and follows you everywhere. Whatever surface you pick tomorrow morning, your clone shows up already knowing you.
FAQ
Claude Cowork vs Claude Code — really, which one?
Same agentic engine. Cowork = pre-configured sandbox, visual, aimed at documents/admin/research ("the work around the work"). Claude Code = folders, terminals, versioning, automation. Non-dev doing office work: Cowork. Anyone building a repeatable system on their own files: Claude Code.
Do the Office add-ins cost extra?
No — included from Pro up. Install from Microsoft AppSource inside each app.
Why do I hit usage limits so fast?
One shared pool (5-hour window + weekly cap) across chat, Code and Cowork — and long sessions re-read their whole history every turn. Shorter, cleaner, separate sessions stretch the same quota dramatically.
Is Claude in Chrome safe to let loose?
It's beta, and Anthropic is candid about it: confirmation on risky actions, finance/adult/piracy sites blocked, injection defenses published — and their own advice is to keep it away from payments and passwords for now.
Which surface should a complete beginner start with?
Chat for a week, then one Cowork task on real files, then the Claude Code guide when you're ready to build a system instead of asking questions.