Anthropic merged Claude chat and Claude Cowork into one conversation on September 16, 2026. The product choice mostly made itself. The choice that still matters is how you brief the work.
Say a marketing coordinator has a Friday afternoon task: pull last month’s campaign numbers from four spreadsheets, compare them to targets, and draft a summary for Monday’s leadership meeting. She opens a chat window, pastes in a question, gets a decent paragraph back, and then has to go find the four spreadsheets herself anyway. The AI helped. It didn’t actually do the task.
That gap, between AI that answers a question and AI that finishes a piece of work, used to be a real product decision. You’d pick between a quick chat, an assistant working inside the app you were already using, or a standing agent you set up separately for anything with genuinely multiple steps. On September 16, 2026, Anthropic collapsed two of those into one conversation: Claude Cowork and Claude chat now live in the same window, and Claude itself decides whether a request needs a quick answer or a multi-step delegated task.[2] “No mode to choose,” as the announcement put it.
So the old question, which product should I open, mostly answers itself now. The question that actually matters hasn’t gone anywhere: how much of this should I hand off, and how much should I stay close to? That’s a decision about the work, not the software, and it’s the one this guide is actually about.
Chat is still the right mode for anything where the output is a thought, not a deliverable. Thinking through a tricky email before sending it. Getting a second opinion on an argument. Asking what a term means, or whether an approach makes sense before committing an afternoon to it. Nothing gets built and nothing gets touched on your computer. You’re thinking out loud with something that talks back.
If you’d be fine with the answer disappearing the moment you read it, it belongs in chat.
A few examples that pass that test:
None of these need a file touched, a connector, or a plan you have to review before it runs. That’s precisely why chat is fast: there’s nothing to approve, because nothing is happening outside the conversation itself.
The next mode is building something concrete while staying close to it the whole time: a document taking shape in Claude Docs, a first-draft deck in Claude Slides, a visual in Claude Design, or a quick interactive tool as an Artifact. You’re not asking for a finished project delivered later. You’re watching it get built, one section at a time, and correcting course as it goes.
This is the mode most people underuse, mostly because it doesn’t feel as impressive as “delegate an agent.” For a large share of real work though, a slide outline, a one-page brief, a chart built from numbers you paste in, this is genuinely the right amount of AI. You get a usable output and you never lose sight of what it’s doing.
| If the work is… | Reach for… |
|---|---|
| A question you want thought through out loud, no file involved | Chat |
| A single document, deck, or visual you want to watch take shape and adjust as you go | Working inside an app (Docs, Slides, Design, Artifacts) |
Both stay inside one conversation now. The difference is whether you’re producing an actual deliverable or just thinking through something.
The third mode used to be a separate product: describing an outcome and letting Claude work out the how, across multiple files, tools and steps, then checking the result afterward instead of watching it happen. Anthropic’s own framing for this mode is direct: “say what, not how.”[1] You’re not writing a set of instructions. You’re describing what done looks like.
Anthropic itself points to four kinds of work as the reason this mode exists at all: scheduled reporting that would otherwise eat a recurring afternoon, reconciling data across spreadsheets and flagging what doesn’t line up, bulk review of a pile of documents or contracts, and research synthesis, pulling scattered notes or calls into one usable summary.[1] Go back to the Friday-afternoon campaign summary from the opening: four spreadsheets, a comparison, a draft for Monday. That’s not a chat question, and it’s too many sources for a single document you’d want to watch get typed line by line. It’s exactly the shape of task this mode exists for.
The genuine trade you’re making by delegating is oversight for time. You’re not watching each step, so the checking happens differently: you review the plan before it runs, and you review the result after, rather than correcting every sentence as it’s written. That’s a real shift in how much trust you’re extending, worth being honest about rather than treating delegation as free.
Describe what done looks like, not the steps to get there, or you’re just writing a longer chat message.
Because this mode touches real files and can take real actions, it’s worth being explicit about what stays yours regardless of how good the delegated draft turns out.
| What the agent does | What you still own | How it gets checked |
|---|---|---|
| Pulls the four spreadsheets, reconciles the numbers against targets, drafts the summary | Deciding which variance is actually worth flagging to leadership Monday | You review the plan before it runs, and read the draft before it goes anywhere, every time |
| Reviews a stack of contracts for one named clause | The judgment call on what to actually do about what it finds | Spot-checked against the original documents, not taken on faith |
Neither task disappears from your plate. What changes is whether you’re doing the first pass yourself or reviewing someone else’s.
Put the three modes side by side and the actual decision gets a lot less abstract. None of this is about which mode is better. It’s about matching the mode to what the task actually needs.
| Task | Best mode | Why |
|---|---|---|
| Deciding how to phrase a difficult email | Chat | Nothing needs to be built or touched, just thought through |
| Turning rough notes into a one-page brief | In-app (Docs) | A single deliverable you’ll want to watch and adjust as it forms |
| Reconciling four spreadsheets against targets and drafting a summary | Delegate | Multiple sources, multiple steps, a clear “done” state you can check afterward |
| Reviewing 40 vendor contracts for one specific clause | Delegate | Repetitive and multi-file, exactly the “bulk review” case Anthropic points to |
| Getting a second opinion on a strategy before a meeting | Chat | You want a sounding board, not an artifact |
A working example, not an exhaustive list. The test that generalizes: does this need a sounding board, a document to watch take shape, or an outcome you can define and check later?
The pattern across every row is the same question asked three different ways: does the work need thinking, does it need building while you watch, or does it need doing without you watching. Once that’s answered, the mode isn’t really a choice anymore either.
The actual expensive mistake isn’t reaching for the wrong mode. It’s reaching for the heaviest one out of habit, delegating a two-minute task the same way you’d delegate a genuinely multi-step one, because delegation feels more advanced.
Delegated work costs more than a quick chat message, both in the oversight you owe it and, on paid plans, in the usage it draws on; higher-usage plan tiers exist specifically because heavier, more autonomous work uses more of a plan’s allowance than a quick question does.[3] None of that makes delegation a bad choice. It makes it the wrong choice for a task a quick chat message would have finished in fifteen seconds.
Match the mode to the task, not to how advanced the tool feels.
If you’re introducing this to a small team, the mistake to avoid is teaching the three modes as a ladder everyone should climb. They’re not a maturity level. Someone doing mostly quick research questions has no real reason to spend a week learning to delegate multi-step work, and someone drowning in recurring spreadsheet reconciliation shouldn’t be told to just get better at chat prompts instead. Ask what kind of work each person actually does before deciding which mode is worth teaching first, and our plain-English guide to what Claude Cowork actually is is a reasonable second step once someone’s spending real time on multi-step tasks.
A few signs a team got the introduction wrong:
Two surfaces sit outside this three-mode picture entirely and are worth naming so they don’t get lumped in by mistake. Claude Code is a separate, specialised surface for actually writing and shipping software, not a mode you’d reach for on a marketing brief. Claude in Chrome pairs with the delegate mode to act on live websites rather than local files, useful when the multi-step work involves a browser, not a spreadsheet. Neither replaces the chat-app-delegate decision above. They’re additional surfaces for specific kinds of work, and most non-technical professionals will rarely need either one.
The other honest caution: none of this replaces judgment on anything that actually matters. A delegated draft of a leadership summary still needs a human deciding what’s worth saying in that room. A quick chat answer to a legal or compliance question still needs someone who actually knows the answer checking it. The mode changes how the work gets done. It doesn’t change who’s responsible for whether it’s right. Once delegating multi-step work is a regular part of someone’s week, our guides on turning Claude Cowork into a real operator rather than a smarter chatbot and a real Claude Cowork setup one coach actually handed over go further than this decision layer does.
Ask what the output actually needs to be. A thought you want talked through, use chat. A single document or deck you want to watch take shape, work inside the app. Multiple files, multiple steps, and a result you’re comfortable checking after the fact rather than watching happen, delegate. Since Anthropic merged Cowork and chat into one conversation in September 2026, this is less about opening a different product and more about how you brief the request.
Anthropic itself names four categories: scheduled reporting, reconciling data across spreadsheets, bulk review of documents or contracts, and research synthesis. Under the hood, this mode can split a complex task into smaller pieces and work on several in parallel, which is part of why it fits multi-step work better than a single chat message does, and why the four categories above share one trait: more than one source or more than one step, with a result you can define in advance.
Not a waste exactly, more like using more tool than the job needs. A single email is usually a fifteen-second chat message. Delegating it doesn’t produce a worse result, it just adds review overhead, a plan to check, a result to read, for a task that never had multiple steps to begin with. Save delegation for work that’s genuinely multi-step or multi-source.
No, and this is one of the more common assumptions worth correcting directly. The output quality on a single, well-scoped question is usually identical whether you ask it in chat or hand it off as a delegated task. What changes with the more autonomous mode is how much of the work happens without you watching, not how good any individual answer is.
Start from the work people already do, not the feature list. A finance analyst who reconciles spreadsheets weekly should learn to delegate that one task first. A salesperson prepping for calls mostly needs the quick-chat mode for thinking through an approach. Teaching all three modes to everyone in one session usually produces people who can describe the features but can’t name a task they actually use any of them for.
The September 16, 2026 Claude Cowork and chat merger details in this guide come from Anthropic’s own announcement (claude.com/blog/cowork-is-now-claude) and product page (claude.com/product/cowork), both fetched directly at write time rather than relied on from memory, since Claude’s product structure changes fast enough that a brief written even a week earlier can go stale. No specific usage-limit numbers are cited, since Anthropic doesn’t publish a fixed figure and any number found in a search risked being outdated by publish.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.