Chat, Projects, Cowork or Code: picking the right mode

Claude runs in four modes and they are not interchangeable. The differences that matter in practice are how much they consume and how much of your attention they need.

The four

Chat. You ask, it answers, you decide what happens next. Cheapest by a wide margin. Right for anything whose output is a paragraph, an explanation, a decision, or a first draft you will edit yourself.

Projects. Chat with a persistent body of context attached across conversations. Right when the same background matters every time and you are tired of re-pasting it. Wrong as a filing cabinet: everything in a project is in the bundle on every turn.

Cowork. An agent working across your files and connected tools, running multi-step tasks while you do something else. Right for work with several steps and a produced artefact at the end. It consumes noticeably more than chat, because it is doing noticeably more.

Claude Code. The terminal agent, for people shipping software. Highest consumption of the four when run loosely, because an agentic loop reads a lot.

Once you have picked a mode, two orthogonal choices remain: the surface you run it on — voice, phone, desktop — and, in chat, whether the answer should become an artifact you can use.

How to choose

Ask two questions in this order.

Does the answer need to become a change in the world — a file written, a system updated, a document produced? If it does not, use chat. This single question settles most cases, and getting it wrong in the expensive direction is the most common way people burn an allowance.

If yes: does it touch a codebase? Code if so, Cowork if not.

Projects is not really a fifth option; it is a modifier on chat for when the context is stable and reused.

The reverse mistake

Using chat for work that genuinely has many steps, and then doing the orchestration yourself across thirty turns. That costs more than the agent would have, takes longer, and the conversation gets worse as it fills.

If you find yourself pasting output from one message into the next as input, you are hand-running an agent loop. Use one.

Try this

For the next three tasks, write down which mode you reached for and why, before starting. Then afterwards, ask whether a cheaper mode would have produced the same result. Most people find at least one of the three should have been a chat.

What goes wrong

Defaulting to the most capable surface. Capability is not free, and the allowance is shared.

Staying in chat out of habit for genuinely multi-step work, then spending an hour being the glue between turns.

Treating Projects as storage. Everything in it is re-sent every turn. It behaves like context, and nothing like a drive.

Switching mid-task because the current one is struggling. Usually the task description is the problem and the surface is fine. Restating the goal well beats escalating.

How to check it worked

Look at what the expensive session actually produced. If the artefact is a paragraph or a decision, that was a chat. If it is a set of files that now exist and did not before, the agent earned its cost.