What you will be able to do
- Choose between web search, extended thinking and Research for a given information need
- Run the Description–Discernment loop on research output, including drawing on internal sources
- Treat Claude-generated plans and process changes as inputs to a human decision
- Describe how Cowork runs a whole task while pausing for approval where you require it
1.Match the information need to the right capability
Claude offers three capabilities for finding information and reasoning about it, and research work goes better when you pick deliberately. With Research on, Claude works agentically. It runs several searches that build on each other, explores different angles of the question, and returns a report with easy-to-check citations. Research requires web search to be turned on. It is available on paid plans, and because it retrieves many sources it can use up your usage limits faster than a normal conversation.
For planning work, the two can be combined: extended thinking lets Claude plan its approach carefully, and Research does the broad information gathering. The help center gives researching emerging technologies for a business proposal as an example.
| Capability | Best for | Example from the help center |
|---|---|---|
| Web search | Straightforward factual queries answerable with one or two tool calls | Getting info about a specific company; recent news headlines |
| Extended thinking | Complex reasoning that doesn't need recent web info | Analysing a question from multiple perspectives |
| Research | Comprehensive gathering: five or more tool calls over 1-3 minutes, synthesised into a report | Compare business competitors; update outdated internal documents with info from the web |
| Extended thinking + Research | Planning the approach, then gathering comprehensively | Researching emerging technologies for a business proposal |
Checkpoint 1 of 6· Check yourself
A planner wants Claude to work out its approach first and then gather information from many sources for a business proposal. What does the help center recommend?
Combining the two lets Claude plan carefully and gather comprehensively. Research cannot work with web search turned off.
“Combining extended thinking with research allows Claude to both plan its approach thoughtfully and execute comprehensive information gathering”Source: support.claude.com
2.The Description–Discernment loop in research
Picking a capability is only the start. The AI Fluency nonprofit course shows research as a loop between two competencies. Description means giving context, not asking a broad question: explain who you are, who you serve, and what you need to know, including time frame and geographic scope. Discernment means reviewing the answer critically: flag specific claims for verification, notice gaps, and question how recent the information is. The exercise asks you to find at least two claims that need verification and to note any missing perspectives. Your first prompt rarely gets you everything, so what you learn from each review shapes a sharper follow-up.
A useful verification habit is to ask Claude to track down the original source for a key claim and compare how accurately its summary represented it.
Description also covers *which* sources Claude draws on. Research runs across your internal context, such as Gmail, Google Calendar and Google Docs when they are connected, as well as the web. When internal knowledge matters, the help center tells you to steer Claude to "Pull relevant context from [relevant internal knowledge source]." Discernment applies to the inputs too. If every source Claude was given comes from one interested party, the result can only reflect that view. Noticing missing perspectives is part of the job.
Checkpoint 2 of 6· Exam question
Which approach best leverages Claude for prioritizing customer feedback?
Correct answer: C — Define clear prioritization criteria (e.g., impact, effort, customer segment), then ask Claude to categorize each feedback item and assign a priority score or rank based on that rubric.
- A. Incorrect. Providing only a broad instruction without a scoring rubric, impact categories, or customer segments leaves the prioritization arbitrary and inconsistent. Claude requires clear guidelines to produce reliable, repeatable outputs aligned with business goals.
- B. Incorrect. While manual categorization into broad themes is a useful first step, delegating only the summarization to Claude misses the opportunity to scale and standardize the prioritization process. Without defined prioritization criteria, the resulting insights may still lack consistent, actionable guidance.
- C. Correct. Defining clear prioritization criteria (impact, effort, customer segment) and using Claude to categorize each feedback item against that rubric ensures a structured, objective, and repeatable prioritization. This approach produces reliable, scalable outputs that directly inform planning.
- D. Incorrect. Fully automating strategic decision-making by having Claude directly decide next quarter's features and provide engineering estimates without human review introduces significant risk of unrealistic or misaligned outcomes. Such decisions require product context and human validation, as LLMs cannot replace the nuanced judgment needed for roadmap planning.
Checkpoint 3 of 6· Check yourself
Claude returns a policy summary for an evidence review. Which response applies Discernment as the AI Fluency course describes it?
Citations make checking easier but don't replace it. The course frames Discernment as flagging claims, noticing gaps and questioning recency.
“Flag specific claims for verification, notice gaps, and question recency of information”Source: academy.claude.com
3.Plans from Claude are inputs; the decision stays yours
Research gets you oriented. Planning turns what you found into options, and Claude can draft several credible plans quickly. The nonprofit course is direct about where that stops: AI speeds up research, but you are still the decision-maker, and you apply professional judgment to every deliverable you own. That matches the outer loop from the 4D Framework. Delegation keeps judgment with you, and Diligence means you own the result, whatever role AI played.
In practice, that means you don't settle a choice between plans by asking Claude to pick. You compare each plan against the success criteria you set before the work started and weigh things that only the people involved know. Then a person decides and can explain why. The same applies when Claude proposes a process change. The proposal is a well-argued starting point, and you test it before acting on it.
Checkpoint 4 of 6· Exam question
A manufacturing engineer has a detailed transcript of a production process and wants to convert it into a Standard Operating Procedure (SOP) using Claude. Which approach best leverages Claude's capabilities to create an accurate, complete, and company-standard SOP?
Correct answer: B — Ask Claude to extract steps from the transcript, then iteratively add safety notes, tools, and quality checks, and format as a company SOP template.
- A. Incorrect. Providing the transcript without specifying details relies entirely on Claude's inference, which may overlook critical safety notes, tools, or quality checks necessary for a complete SOP. This approach also lacks iterative refinement to ensure the output matches company-specific formatting and content requirements, increasing the risk of omissions or inaccuracies.
- B. Correct. This method leverages Claude's capabilities by breaking the task into logical steps: extracting process steps, then iteratively adding safety notes, tools, and quality checks, and finally formatting into a company SOP template. This iterative process ensures the SOP is accurate, complete, and aligned with company standards.
- C. Incorrect. Limiting Claude to only grammar and style checks underutilizes its ability to analyze the transcript, suggest organizational improvements, and ensure procedural completeness. Manual rewriting and structuring risk introducing errors or missing details that Claude could otherwise catch and enhance.
- D. Incorrect. Starting with a generic SOP template and inserting details from the transcript can lead to misalignment between the template's structure and the actual process, requiring extensive manual review to correct. This approach fails to capitalize on Claude's capacity to directly generate a tailored, accurate SOP from the source material.
Checkpoint 5 of 6· Check yourself
According to the AI Fluency nonprofit course, what is AI's role in research that feeds a decision?
The course says AI accelerates research but doesn't replace expertise. The person remains the decision-maker.
“who must apply professional judgment to every deliverable you own”Source: academy.claude.com
4.Cowork: taking on the whole task, with approval where you want it
Once a process step is delegated, Claude Cowork is where Claude carries it out from start to finish. Anthropic's guidance puts the line this way: use chat for anything that fits in a few exchanges, and use Cowork when you need a deliverable or a multi-step task that touches several files or apps. The flow is to describe the task, review Claude's approach, then let it run. While it works, progress indicators show each step, and Claude shows its reasoning so you can follow along. You can step in when it matters or let Claude run on its own.
Approval checkpoints are set through modes that control when Claude asks permission before acting, for example before using a connector. On Team and Enterprise plans, an admin controls whether automatic approval is available, and the organisation may require approval for each task that uses write-capable connector tools.
| Mode | Behaviour |
|---|---|
| Manual (Manually approve) | Claude pauses and asks for approval; you choose Allow or Deny |
| Auto (Automatically approve) | Claude decides; read-only tools set to Always allow are approved |
| Skip | Approved |
Checkpoint 6 of 6· Check yourself
A team wants Cowork to compile a recurring report through connectors, but someone must approve every action before it happens. Which mode fits?
In Manual mode, Claude pauses at each action and you choose Allow or Deny. In Auto, Claude makes some of those calls itself.
“Claude pauses and asks for approval for actions. You review each request and choose Allow or Deny.”Source: support.claude.com
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Output from Research comes with citations, so it is ready to use as is.Why is that wrong?
Citations make claims easy to check, not checked already. Discernment still means flagging claims, spotting gaps and questioning recency.
Covered in The Description–Discernment loop in research
2.When Claude produces competing plans, the fastest correct step is to ask Claude which one is better and adopt it.Why is that wrong?
A person stays the decision-maker. They compare the plans against criteria they set and apply professional judgment to the deliverable they own.
Covered in Plans from Claude are inputs; the decision stays yours
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
“Web search is best for straightforward, factual queries that can be answered with one or two tool calls”
↩︎ Match the information need to the right capability“compare business competitors, update outdated internal documents with info from the web”
↩︎ Match the information need to the right capability“Research is optimal for comprehensive information gathering requiring five or more tool calls over 1-3 minutes”
↩︎ Prediction“Combining extended thinking with research allows Claude to both plan its approach thoughtfully and execute comprehensive information gathering”
↩︎ Checkpoint - 2.
“Important: You must have web search turned on for research to function.”
↩︎ Match the information need to the right capability“research sessions can use up your limits faster due to Claude retrieving multiple sources and providing comprehensive responses.”
↩︎ Match the information need to the right capability“across your internal context (such as Gmail, Google Calendar, and Google Docs when connected) and the web”
↩︎ The Description–Discernment loop in research“Pull relevant context from [relevant internal knowledge source]”
↩︎ The Description–Discernment loop in research - 3.
“explain who you are, who you serve, and what you specifically need to know”
↩︎ The Description–Discernment loop in research“Note any missing perspectives relevant to your communities”
↩︎ The Description–Discernment loop in research“Ask AI to track down the original source for one key claim in its summary and compare how accurately it was represented”
↩︎ The Description–Discernment loop in research“who must apply professional judgment to every deliverable you own”
↩︎ Plans from Claude are inputs; the decision stays yours“Flag specific claims for verification, notice gaps, and question recency of information”
↩︎ Exam trap 1“who must apply professional judgment to every deliverable you own”
↩︎ Exam trap 2“Flag specific claims for verification, notice gaps, and question recency of information”
↩︎ Checkpoint - 4.
“Diligence means owning the final result fully”
↩︎ Plans from Claude are inputs; the decision stays yours - 5.
“so you can steer when it matters, or let Claude run independently”
↩︎ Cowork: taking on the whole task, with approval where you want it“Cowork has three modes that control when Claude asks your permission before taking an action, like using your connectors.”
↩︎ Cowork: taking on the whole task, with approval where you want it“Describe the task you want Claude to complete.”
↩︎ Cowork: taking on the whole task, with approval where you want it“Claude pauses and asks for approval for actions. You review each request and choose Allow or Deny.”
↩︎ Checkpoint - 6.
“Use chat if what you want fits in a few exchanges, like a question, an explanation, a brainstorm, or a gut check.”
↩︎ Cowork: taking on the whole task, with approval where you want it