What you will be able to do
- Explain why the same prompting habits produce differently shaped prompts depending on the task type
- Structure an analysis prompt with a defined output format, permission to express uncertainty, and grounding in the source text
- Write a context-rich research prompt, choose between web search, extended thinking and Research, and plan how to verify the result
Key concept
Task-shaped prompting — Clarity, context and specificity apply to every prompt, but what you have to spell out changes with the job. Analysis needs a defined output shape and room to admit uncertainty. Research needs your situation and a plan for checking the answer. Drafting needs voice and details only you know, and brainstorming needs an explicit request for breadth within stated constraints.
1.Same principles, different prompt shapes
Anthropic's prompting guidance starts with a handful of habits: give clear, explicit instructions, explain the reason behind a constraint, and say exactly what output you want. Its golden rule is to show your prompt to a colleague who knows little about the task. If that colleague would be confused, Claude will be too. These habits don't change between tasks. What changes is the information a new colleague would be missing, and that depends on the job you're handing over.
Ask a new analyst to 'look at these numbers' and they need to know what shape the findings should take, and whether it's all right to say the data isn't enough. Ask them to 'find out about the Seattle funding landscape' and they need to know who you are and why you're asking. A writer needs your voice and the facts only you have. Someone brainstorming needs to hear that you want many options, and what limits those options must respect. This page covers the first two task types, analysis and research. Drafting and brainstorming follow the same logic with different ingredients.
Sources1
2.Analysis: define the output, allow uncertainty, ground it in the text
In an analysis task, you already have the material, such as a report, a dataset or competitor notes, and you want a judgement about it. The two usual failures are output in whatever shape Claude happens to choose, and confident conclusions the data doesn't support. Each has a direct fix in the prompt.
First, specify the structure. If you're comparing several things along the same dimensions, show the format once and ask Claude to fill it in for each item. Anthropic's consistency guide does this for competitor analysis: a SWOT template with word limits on the overview and the strategy line.
<competitor>
<name>Rival Inc</name>
<overview>A 50-word summary.</overview>
<swot>
<strengths>- Bullet points</strengths>
<weaknesses>- Bullet points</weaknesses>
<opportunities>- Bullet points</opportunities>
<threats>- Bullet points</threats>
</swot>
<strategy>A 30-word strategic response.</strategy>
</competitor>Second, give Claude a way out. The hallucination guide recommends saying explicitly that 'I don't know' is an acceptable answer. Its example names the role and the focus areas, and it scripts the exact words to use when the report falls short:
As our M&A advisor, analyze this report on the potential acquisition of AcmeCo by ExampleCorp.
<report>
{{REPORT}}
</report>
Focus on financial projections, integration risks, and regulatory hurdles. If you're unsure about any aspect or if the report lacks necessary information, say "I don't have enough information to confidently assess this."Third, for long source documents (over roughly 20k tokens), ask Claude to pull out word-for-word quotes before it analyses anything, and to base the analysis only on those quotes. That ties each conclusion to the text you supplied. Finally, when a judgement takes several reasoning steps, you can ask Claude to work step by step before giving a verdict. Where extended thinking is available, it's generally the better choice. Asking for visible step-by-step reasoning still helps when you want to review the reasoning yourself or make sure particular factors are considered.
A team wants Claude to research an open-ended market question, tracking multiple competing explanations and its confidence in each as it gathers evidence over many searches. What approach best supports this research task?
Correct answer: A — Develop competing hypotheses, track confidence in notes, and self-critique while searching
- A. Correct. For complex, open-ended research, Claude should develop competing hypotheses, track confidence in progress notes, and self-critique regularly to work through evidence methodically.
- B. Incorrect. Committing to one explanation immediately biases the research toward confirming a single theory instead of weighing competing hypotheses against gathered evidence.
- C. Incorrect. A single unverified search skips the source verification that complex research questions need before drawing conclusions.
- D. Incorrect. Discarding intermediate findings loses the calibration and transparency that a persisted research notes file is meant to provide across a long research task.
3.Research: supply your context, pick the right capability, plan to verify
In a research task, Claude gathers information you don't yet have. Anthropic Academy's research walkthrough follows Maria, an executive director expanding a housing nonprofit from Portland to Seattle. The main lesson is that a broad question gets a generic answer. Her prompt improves when she states who she is, whom her organisation serves, and what she needs to know. The accompanying exercise lists the parts of a good research prompt: the specific policy or subject, your organisation's context, what you need to learn (impact, funding, compliance, opportunities), and a time frame or geographic scope.
How you word the prompt is half the choice. The other half is which capability you use. The Claude help centre separates three options by how much searching and reasoning the task needs:
| Capability | Best suited to | Example from the help article |
|---|---|---|
| Web search | Straightforward factual queries answerable with one or two tool calls, or when you want to steer what gets searched | Recent news headlines; who won last night's game |
| Extended thinking | Complex reasoning that doesn't need recent information from the web | Optimising an algorithm; analysing bioethics from multiple perspectives |
| Research | Comprehensive gathering of five or more tool calls over 1-3 minutes, synthesised into an in-depth report | Comparing business competitors; updating outdated internal documents |
| Extended thinking + Research | Planning the approach carefully, then gathering information comprehensively, with citations in a longer report | Researching emerging technologies for a business proposal |
The research prompt isn't finished when the answer comes back. The Academy course pairs Description, which is what you tell Claude, with Discernment, which is how critically you judge what comes back. In research, Discernment means marking claims that need checking, noticing missing perspectives, and questioning whether information is current. The loop is iterative by design: the first prompt rarely gives you everything, and what you spot in the answer shapes a sharper follow-up. The exercise's stretch goal fits here as well. Ask Claude to find the original source for a key claim, then compare how accurately the summary represented it.
Geographic and time scope (which city or state, and whether you mean current rules or upcoming ones), the organisation's context (for example, a housing nonprofit serving low-income renters), and what they actually need to know (compliance obligations, effects on beneficiaries, or advocacy opportunities). Without these, Claude can only give a generic overview.
A researcher asks Claude to answer a factual question that could be verified from several independent sources before being treated as reliable. Which instruction best supports getting a trustworthy research answer?
Correct answer: A — Ask Claude to cross-check the claim against several sources before calling it settled
- A. Correct. Encouraging source verification, checking a claim across multiple sources, is a recommended practice for getting reliable results on research tasks.
- B. Incorrect. Relying on the first source found skips the verification step needed to establish that a research finding is reliable.
- C. Incorrect. Confident phrasing unrelated to actual verification can misrepresent how well-supported the claim really is.
- D. Incorrect. Hiding which sources were consulted removes transparency and makes it harder to judge how well the claim was actually verified.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.An analysis prompt should demand a definitive conclusion every time, because allowing 'I don't know' makes the answer less useful.Why is that wrong?
Explicitly allowing Claude to say the data is insufficient makes an analysis more reliable, not weaker. A forced verdict on thin evidence is where false information comes from.
Covered in Analysis: define the output, allow uncertainty, ground it in the text
2.A short, generic research question gives the most objective results, because adding your own situation biases the answer.Why is that wrong?
Context is what makes research output relevant and actionable. Say who you are, whom you serve, and exactly what you need to know.
Covered in Research: supply your context, pick the right capability, plan to verify
3.Any question that needs current information should go to the Research feature.Why is that wrong?
Web search fits a straightforward factual lookup. Research is for comprehensive, multi-source gathering that ends in an in-depth report.
Covered in Research: supply your context, pick the right capability, plan to verify
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practicesOfficial docs
“Be specific about the desired output format and constraints.”
↩︎ Same principles, different prompt shapes“Think of Claude as a brilliant but new employee who lacks context on your norms and workflows.”
↩︎ Same principles, different prompt shapes“Think of Claude as a brilliant but new employee who lacks context on your norms and workflows.”
↩︎ Key concept - 2.https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/increase-consistencyOfficial docs
“Output following this example format:”
↩︎ Analysis: define the output, allow uncertainty, ground it in the text - 3.https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinationsOfficial docs
“ask Claude to extract word-for-word quotes first before performing its task”
↩︎ Analysis: define the output, allow uncertainty, ground it in the text“Explicitly give Claude permission to admit uncertainty. This simple technique can drastically reduce false information.”
↩︎ Exam trap 1 - 4.https://claude.com/blog/best-practices-for-prompt-engineeringSecondary source
“This technique helps with complex analytical tasks that benefit from structured thinking.”
↩︎ Analysis: define the output, allow uncertainty, ground it in the text - 5.
“Flag specific claims for verification, notice gaps, and question recency of information”
↩︎ Research: supply your context, pick the right capability, plan to verify“Your first prompt rarely gives you everything you need.”
↩︎ Research: supply your context, pick the right capability, plan to verify“Don't just ask broad questions—explain who you are, who you serve, and what you specifically need to know.”
↩︎ Exam trap 2 - 6.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”
↩︎ Research: supply your context, pick the right capability, plan to verify“Web search is best for straightforward, factual queries that can be answered with one or two tool calls”
↩︎ Exam trap 3