What you will be able to do
- Choose between web search, extended thinking and Research based on the shape of a research question
- Explain what Claude's Research feature does, what it can draw on, and what it costs in usage
- Write a research prompt that gives Claude the context it needs
- Apply Discernment to research output before it becomes part of a deliverable
Key concept
The Description-Discernment loop — Research with Claude is a cycle. You describe your situation and need in context-rich terms, critically evaluate what comes back, and use what that evaluation reveals to ask a sharper follow-up. You remain the decision-maker throughout.
1.Match the question to web search, extended thinking or Research
Using Claude for research starts before the first prompt. You first decide how much searching the question actually needs. Claude has three capabilities that overlap, and Anthropic's guidance separates them by the shape of the task, not by how important it is. A quick factual check, such as one company's details or a recent headline, takes one or two searches. A hard reasoning problem that doesn't depend on fresh information needs Claude to think longer, not to search. A question that needs many sources combined into a report is what Research is for.
| Capability | Best for | Example from the guidance |
|---|---|---|
| Web search | Straightforward factual queries that can be answered with one or two tool calls | Getting info about a specific company; retrieving recent news headlines |
| Extended thinking | Complex reasoning tasks that don't require recent info from the web | Analyzing the bioethics of organ donation from multiple perspectives |
| Research | Comprehensive information gathering: five or more tool calls over 1-3 minutes, synthesized into an in-depth report | Comparing business competitors; updating outdated internal documents with info from the web |
| Extended thinking + Research | Planning the approach thoughtfully, then gathering information comprehensively | Researching emerging technologies for a business proposal |
Two details in that guidance matter in practice. First, web search is also the choice when you want to steer more closely: you tell Claude roughly what to search for, and it runs the searches and analyzes the results. Second, combining extended thinking with Research is the option for work that needs a plan. Claude thinks through how to approach the question before it starts gathering, which suits a business proposal or a review of several papers.
Sources1
2.What Research does, and what it can draw on
Research doesn't fetch once and answer. Claude works agentically: each search builds on the last, and Claude decides what to investigate next. Anthropic's engineering write-up explains why this fits research. The steps can't be predicted in advance, and human researchers work the same way, following leads as they turn up. The write-up contrasts this with static retrieval, which pulls the chunks most similar to the query. Research instead runs a multi-step search that adapts to what it finds. Internally, an agent plans the research process and then launches parallel agents that search at the same time.
Research doesn't search only the public web. It also covers your internal context, such as Gmail, Google Calendar and Google Docs when those are connected. If the answer should come from a specific internal source, say so in the prompt. The help article's wording is to prompt Claude to pull relevant context from that named source. Don't assume the web results will cover it.
Some practical limits apply. Research is available on paid plans (Pro, Max, Team or Enterprise) and needs web search turned on. It counts against the same limits as ordinary conversations but uses them up faster, because it retrieves many sources. The engineering team reports the same pattern at a larger scale: agents use about 4× more tokens than chats and multi-agent systems about 15× more, so the task has to be valuable enough to justify the cost.
3.Description: give Claude the context a researcher would need
Anthropic's AI Fluency course shows research through one scenario. Maria, an executive director, is expanding her housing nonprofit from Portland to Seattle. She uses AI to research policy, funding and compliance. The first lesson is Description. A broad question gets a broad answer, so a useful research prompt says who you are, who you serve, and exactly what you need to know.
The course exercise turns this into a checklist. A research prompt should name the specific policy or topic, give your organisation's context and why the topic matters to it, and state what you need to find out: impact, funding implications, compliance requirements or opportunities. It should also set a time frame or geographic scope. Treat the first answer as a draft of the research, not the result. What you learn from evaluating it goes into the next, more targeted question.
A stronger version names who is asking and why, what exactly they need, and the scope. For example: 'I lead a housing nonprofit that is expanding its services from Portland to Seattle. Summarise the Seattle housing policies that would affect our beneficiaries, what funding they open up for an organisation like ours, and what compliance requirements we would take on. Focus on rules currently in force in Seattle.'
Sources4
4.Discernment: verify before research becomes a deliverable
Research returns its findings with citations that are easy to check. Anthropic's pipeline even ends with a dedicated step that attributes each claim to its source. Citations make checking possible. They don't make it unnecessary. The AI Fluency course treats Discernment as required, especially when accuracy matters.
The course's Discernment exercise makes this concrete. Pick out at least two claims that need verification. Note any missing perspectives that matter to the people you serve. Flag anything that looks outdated or too general. The stretch goal is the most revealing: ask the AI to find the original source for one key claim, then compare how accurately its summary represented that source. Also look at where the sources came from. A summary built only from material that parties with a stake in the answer wrote about themselves is likely to have exactly the gaps this exercise asks you to find.
A market researcher needs Claude to incorporate current, real-world information published this week into a competitive-landscape summary, since this data postdates the model's training. Which capability should the researcher enable?
Correct answer: C — The web search tool, which augments Claude's knowledge with current data from across the web
- A. Prompt caching lowers cost and latency for repeated context; it does not fetch new external information.
- B. The memory tool persists information across sessions but does not retrieve fresh external web content.
- C. Correct. The web search tool augments Claude's comprehensive knowledge with current, real-world data from across the web, which is needed for this week's information.
- D. Structured outputs controls response formatting, not access to current external information.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Research is the best choice for any question you want answered well.Why is that wrong?
Research is for comprehensive gathering over many tool calls. A straightforward factual query is better served by web search, which answers it in one or two tool calls, and Research uses up plan limits faster.
Covered in Match the question to web search, extended thinking or Research
2.Research output that comes with citations can go into a deliverable without further checking.Why is that wrong?
Citations make checking easy, but Discernment still means flagging specific claims for verification, noticing gaps and questioning how recent the information is before you rely on it.
Covered in Discernment: verify before research becomes a deliverable
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://support.claude.com/en/articles/11095361-when-should-i-use-web-search-extended-thinking-and-researchOfficial docs
“Research is optimal for comprehensive information gathering requiring five or more tool calls over 1-3 minutes”
↩︎ Match the question to web search, extended thinking or Research“Extended thinking shines when tackling complex reasoning tasks that don't require recent info from the web”
↩︎ Match the question to web search, extended thinking or Research“Combining extended thinking with research allows Claude to both plan its approach thoughtfully and execute comprehensive information gathering”
↩︎ Match the question to web search, extended thinking or Research“It's also helpful if you want to steer Claude more”
↩︎ Match the question to web search, extended thinking or Research“Web search is best for straightforward, factual queries that can be answered with one or two tool calls”
↩︎ Exam trap 1 - 2.
“Claude operates agentically, conducting multiple searches that build on each other while determining exactly what to investigate next.”
↩︎ What Research does, and what it can draw on“Claude will kick off the research process across your internal context (such as Gmail, Google Calendar, and Google Docs when connected) and the web.”
↩︎ What Research does, and what it can draw on“Pull relevant context from [relevant internal knowledge source].”
↩︎ What Research does, and what it can draw on“Research is available for users with paid Claude plans (Pro, Max, Team, or Enterprise)”
↩︎ What Research does, and what it can draw on“You must have web search turned on for research to function.”
↩︎ What Research does, and what it can draw on“research sessions can use up your limits faster due to Claude retrieving multiple sources and providing comprehensive responses.”
↩︎ What Research does, and what it can draw on“complete with easy-to-check citations so you can trust Claude's findings”
↩︎ Discernment: verify before research becomes a deliverable - 3.
“When people conduct research, they tend to continuously update their approach based on discoveries, following leads that emerge during investigation.”
↩︎ What Research does, and what it can draw on“our architecture uses a multi-step search that dynamically finds relevant information, adapts to new findings, and analyzes results to formulate high-quality answers.”
↩︎ What Research does, and what it can draw on“an agent that plans a research process based on user queries, and then uses tools to create parallel agents”
↩︎ What Research does, and what it can draw on“multi-agent systems require tasks where the value of the task is high enough to pay for the increased performance.”
↩︎ What Research does, and what it can draw on“This ensures all claims are properly attributed to their sources.”
↩︎ Discernment: verify before research becomes a deliverable - 4.
“Don't just ask broad questions—explain who you are, who you serve, and what you specifically need to know.”
↩︎ Description: give Claude the context a researcher would need“Use what you learn from evaluating each response to craft better, more targeted follow-up questions”
↩︎ Description: give Claude the context a researcher would need“Time frame or geographic scope”
↩︎ Description: give Claude the context a researcher would need“Discernment isn't optional”
↩︎ Discernment: verify before research becomes a deliverable“Note any missing perspectives relevant to your communities”
↩︎ Discernment: verify before research becomes a deliverable“Ask AI to track down the original source for one key claim in its summary and compare how accurately it was represented.”
↩︎ Discernment: verify before research becomes a deliverable“The Description-Discernment loop is iterative: Your first prompt rarely gives you everything you need.”
↩︎ Key concept“You must evaluate AI outputs critically, especially when accuracy matters. Flag specific claims for verification, notice gaps, and question recency of information”
↩︎ Exam trap 2