Claude Chat vs Cowork vs Claude Code: which should you use?

One-line decision tree (quotable)

Use chat to work through an answer, Cowork to delegate a defined deliverable, and Claude Code to build or maintain a project.

My judgement starts with the work you want back. A useful conversation, a finished report and a working application need different ways of organising the task. Choosing by brand or model alone misses that distinction.

There is a naming complication. Anthropic says Chat and Cowork are gradually becoming one Claude experience, starting with Pro and Max, with an Enterprise beta. Some accounts will no longer show separate Chat and Cowork choices. I retain the names here to describe the workflows, rather than promise separate buttons. See the official getting-started guide.

What a "harness" is (Craft framing)

A harness is the working arrangement around the model. It determines how a request becomes steps, which tools are available, where work happens and how you inspect the result. This is my explanatory framing, not an Anthropic product tier.

In Craft I use seven layers: model, harness, connections, skills, product, agency, my role. The model interprets and generates. The harness organises action. Connections provide access. Skills supply a repeatable procedure. The product wraps those pieces in an interface and commercial terms. Agency describes how much action I delegate. My role identifies what I must still decide.

The practical question is simple: what evidence would let me accept the result? For advice, I need sound reasoning. For a report, I need traceable inputs. For software, I need observable behaviour and appropriate checks. That acceptance test should shape the harness I choose.

Craft modes describe my working relationship with AI. In Assist, AI helps me think; I decide. In Augment, AI carries several steps; I review. In Adapt, AI works on files within rules I set; I stay on call. These are editorial descriptions, not vendor capability limits.

Chat

Start with chat when the next useful output is something you can read and respond to: an explanation, a critique, an outline or a set of options. I use it when the judgement is still forming and handing over execution would be premature.

Do not equate chat with text alone. Claude pricing lists file creation, code execution and connected tools on Free. The distinction is therefore the shape of the assignment, rather than a blanket claim that chat cannot make files or use tools.

The Claude chat profile is the starting tracker entry. My default remains Assist when I am exploring a sensitive or ambiguous decision. I can move to Augment once the inputs, output and review criteria are clear.

Cowork

Choose Cowork when you can describe a bounded piece of work and inspect its deliverable. Anthropic's Cowork overview documents direct local file work through Claude Desktop, document production, browser integration with Claude in Chrome and coordination through sub-agents. No terminal is required.

My useful starting prompt names the source folder, the required result and the exclusions. For example: “Use these approved interview notes to draft a findings report. Link each finding to its source. List contradictions separately. Save a new draft for review.” That is an example workflow, not a claim about guaranteed quality.

Cowork is included on paid Claude plans. Surface availability and administrative enablement vary, so consult the current help page before assuming your account has the same interface as a colleague. The Cowork profile and Claude Cowork vs Copilot Cowork comparison cover the wider product decision.

Code

Choose Claude Code when the deliverable belongs to a project that you need to inspect, run and maintain. Its official overview describes reading a codebase, editing files and running commands. It is available through the terminal, desktop app, browser, VS Code and JetBrains integrations.

A terminal is therefore an option, rather than the only way to use Code. The bigger requirement is understanding what a successful project looks like. If you request a small internal tool, someone must check the inputs, calculations, access rules and failure behaviour.

For a non-developer, Code becomes worthwhile when repeatability matters enough to justify that review. Consider a recurring file conversion with explicit rules and sample inputs. A maintained project could make the work easier to repeat. If the requirement changes every week and no one can inspect the result, a bounded Cowork task may be the better starting point.

I would write acceptance examples before commissioning the tool: one ordinary input, one incomplete input and one input that should be rejected. Those examples make the intended behaviour reviewable. They also expose uncertainty before the project becomes harder to change.

See the Claude Code profile for the tracker entry.

Shared vs separate (skills, plugins, connectors)

Shared concepts do not mean every installation is shared automatically. The extensibility overview says account-added connectors, skills and plugins follow your signed-in account, including into Claude Code. Command-line additions stay on that machine.

The Cowork overview is explicit about the reverse direction: Cowork loads what is enabled for your claude.ai account at session start and does not read the CLI's ~/.claude directory. Add a local-only skill or plugin through Customize if you want it available there.

A skill describes a procedure. A connector supplies access. A plugin packages extensions. Check the actual components and permissions when moving a workflow between surfaces. A workflow that assumes a particular local file path needs more than the same name appearing in another interface. See skills vs MCP vs plugins.

For an individual subscription, chat starts Free. Pro includes Claude Code and task delegation; the US page lists $20 monthly or $200 annually, displayed as $17 per month. Usage limits apply, with activity across Claude surfaces and Code drawing from a shared pool. These are US-page figures, not Australian-dollar quotes.

Decision by task, with 6 examples

These are my suggested starting points. The final choice depends on the sources, permissions and reviewer.

Task	Start with	What I would check
Challenge a strategy proposal	Chat	Assumptions and missing evidence
Prepare a report from approved interviews	Cowork	Findings against the notes
Organise a defined folder of working documents	Cowork	Proposed names and destinations
Build a small internal calculator	Code	Calculations, edge cases and access
Turn an unclear request into a usable brief	Chat	Scope and acceptance criteria
Maintain a repeatable file-processing project	Code	Sample outputs and recovery behaviour

For every task, name the stop point. “Draft for my review” is a different delegation from “update the working files”. Agree where uncertainty should be reported and which actions remain yours. More autonomy is useful only when the work becomes easier to verify.

FAQ

Is Claude Code better than Cowork?

Claude Code is the better starting point for building and maintaining a project; Cowork suits a defined knowledge-work deliverable. That is my task-based judgement, not a ranking of intelligence. Compare the acceptance evidence you need: working behaviour for a project, or a traceable report and inspectable files for a delegated assignment.

Do I need the terminal?

No. Cowork does not require a terminal, and Claude Code also has desktop, browser and editor surfaces. Choose the interface you can review confidently. Avoid treating a friendlier interface as proof that the work needs less supervision: you still need to understand the intended changes and recognise a bad result.

Do Cowork and Code share skills?

Account-enabled skills can reach both, but local-only Code files are separate. Cowork does not read your machine's ~/.claude directory. Add those extensions through Customize for account use. Before relying on a shared skill, check its references, tool assumptions and permissions in the surface where you intend to run it.

Which plan should I choose?

Start with Free for chat, then consider a paid plan when you need Cowork or subscription access to Claude Code. Pro is the entry individual subscription shown here. Evaluate representative tasks before upgrading for capacity: the amount of review and the complexity of delegated work affect practical value alongside the fee.

When should a non-developer use Claude Code?

Use it when you need a repeatable project and can arrange competent review of the result. Write examples of correct and incorrect behaviour first. If you only need a one-off document transformation, try Cowork. If the task itself is unclear, use chat to define it before asking either harness to execute.

Craft starts with picking the harness. Practise moving from Assist to Augment and Adapt with a clear review role.

Verification notes

• Official pages describe a gradual Chat/Cowork merger. Names and interface steps can differ by account during rollout.
• Pricing lists Skills on Free, while the skills overview lists paid plans. Free-plan skill availability is therefore disputed in the official documentation Not yet confirmed; this comparison does not rely on it.
• Craft modes, task recommendations and acceptance examples are editorial guidance. No model-version or performance ranking is asserted.