Chat, work, or code mode, choosing the right AI surface

You have the answer. It is sitting in a chat window, the deck is due at four, and so you start copying numbers across by hand and reformatting each one into the template.

That afternoon is the tell. One connection, three surfaces, and you picked the wrong one for the thing you were making. Chat answers questions. Work mode builds documents. Code mode automates.

Underneath all three sits one MCP, the standard that lets an assistant read a data source directly, so you ask your questions where you already work instead of exporting spreadsheets. Most people sort this out by accident around week three. Here is the shortcut.

Chat is for answers and follow-ups

Anything whose output is a sentence.

Claude chat and ChatGPT carry the conversational work: the Monday check, the drop triage, the share question you need settled before a meeting. Cards render inline. Follow-ups keep their context, and the exchange takes the shape of asking a colleague who already knows your setup.

Chat is also where the scope conversation happens. Pin your segment and your dates once and every follow-up inherits them. That inheritance is most of what makes the surface feel like talking to a person rather than filling in a form. It is the default for most people most days, and every example on the prompts page assumes you are here.

Work mode is for deliverables

When the output is a file, this is the surface. A monthly deck, a competitor readout, a formatted report.

The assistant runs the same analyses it would have run in chat, then assembles the results into something you can send. Scope and the as-of date travel onto the slides. That date is the day the data was last complete, printed on the artifact so a lag never reads as a loss.

The tell that you are on the wrong surface is the copying. The moment numbers start moving out of a chat window into a slide template by hand, the task has outgrown chat. The worked example below shows the deck-shaped version of a competitive readout.

Longer jobs belong here too. A readout that touches several analyses, gathers the results and ships one artifact is a long scroll in chat and a single file in work mode. The difference compounds every month you repeat it.

The workflow that does this: Competitor movers, deck-ready →

Code mode is for pipelines

Here the job has to run without you watching it.

Claude Code and the editor surfaces, Cursor and VS Code, run the same tools from a terminal, with results as text instead of rendered cards. That sounds like a downgrade until the task is programmatic: pulling results into a script, wiring analyses into a pipeline you already maintain, batch-checking fifty keywords on a schedule.

Text output is the feature here. A script wants values it can parse. A scheduled job wants a result it can log. Neither wants a rendered card, and forcing one through a chat window is how people decide the whole thing is fiddly.

Engineers on search teams tend to live in this surface already. The connection meets them there without asking them to visit a chat window first. Same data, same checks, their environment.

One connection under all three

Nothing about the surface changes what comes back.

The same connection, the same scope rules, the same refusals apply in a terminal as in a chat bubble. Two teammates on two surfaces argue from the same numbers. That is the quiet condition that lets a mixed team work at all: the analyst in chat, the content lead in work mode, the engineer in code, one source underneath the three of them.

Access matches too. One read-only connection per person, honoured the same way everywhere, so adding a surface never means reopening a permission conversation. A segment you pinned in one place means the same thing when you ask from another, because the scope lives with your account rather than with the window.

Where does setup actually differ?

In the connection step, and nowhere after it.

Each client has its own two-minute path. ChatGPT wants developer mode switched on, the editors take a server entry, and the docs walk each one through. If a tool list looks stale after an update, reconnecting refreshes it. The article on evaluating a connection covers what to check before you add one.

So pick by deliverable, connect once per surface, and let the question decide where you ask it. Start on one surface. Add the second when a task obviously belongs there.

Most people meet this in chat, hit their first deck-shaped task within a couple of weeks, and find code mode the day somebody asks for the same report a third time. Nobody needs all three on day one.

See also: what to check before you add any connection →