Chat, web search, research, extended thinking, Projects and artifacts

A pricing analyst uploads two years of bookings and the rate rules for her region. She wants to understand how a weekend surcharge would interact with loyalty discounts and minimum-rental rules. A colleague tells her research is “the most powerful mode” and that she should always turn it on for important questions. Claude spends several minutes searching the web and returns a report of general articles on rental pricing, with little analysis of her rules.

Everything Claude needed was already in the files. She judged the importance of the question correctly and picked the wrong tool.

What each feature does

The product features of claude.ai solve different things:

Feature	What it does	What it’s for
Chat	Conversation with what you give it and what the model already knows	Drafting, summarizing, analyzing material you paste or upload
Web search	One or a few searches to bring in current information	Specific questions: a recent fact, a piece of news
Research	Several searches that build on each other; decides what to look at next based on what it finds, and produces a report with citations	Broad investigations with many sources
Extended thinking	Claude reasons step by step before answering	Complex problems with the material you already have
Project	A space with instructions and documents that apply to all its conversations	Recurring work on the same topic or team
Artifact	A piece of content beside the chat, to edit and share	Documents, tables or tools you’ll reuse

Research or web search

Both go to the web; the scale is what changes. A question answered with one or two searches (“when does the new labelling rule take effect?”) doesn’t need research. An overview of the premium cat food market in three countries, with brands, launches, prices and labelling rules, with sources leadership can check, does: that’s exactly what research is for.

Research is available on paid plans, needs web search turned on and uses up your usage limit faster than a normal conversation. Its citations are checked like any others (see the evaluation domain). Pasting a long list of questions into a regular chat doesn’t replace it: they’re still separate answers, without the chained search.

Extended thinking: when the hard part is reasoning

In the analyst’s case, the hard part wasn’t finding information but reasoning about how three rules interact on data that was already uploaded. That calls for extended thinking: Claude works through its reasoning before answering (you can expand it to see it), and it pays off in analyses with many moving parts, planning and detailed documents.

The quick rule: if the answer depends on recent outside information, web search or research; if it depends on reasoning well about what you already have, extended thinking.

Projects and artifacts, in one line

• Project: when you repeat the same instructions or upload the same documents in every chat. Not for a one-off question. (How to set one up is in the configuration domain.)
• Artifact: when the output is a piece to edit or share. It doesn’t organize an investigation or replace research. (When it beats inline is in the evaluation domain.)

Common traps

• Research for everything because it’s “the most powerful”: slow, uses more of your limit and searches outside for what was inside.
• A regular chat with twenty pasted questions for a broad investigation.
• An artifact as if it were a way to research. It changes where the answer appears, not how it’s obtained.
• A new Project for a one-off question.
• Web search for an analysis that depends only on your files.

On the exam

Notice where the information is and what the hard part is. “Many sources,” “with citations,” “several countries,” “overview” point to research. “Everything is in the files she uploaded” and “how the rules interact” point to extended thinking. “A recent fact” points to web search. “Repeats the same instructions every time” points to a Project. Rule out the option that picks the most powerful feature without looking at the task.