CertSafari
    Anthropic Claude Certified Associate – Foundations (CCAO-F)· Lessons

    Domain 1 · Lesson 4/30

    Prompting by task type: analysis, research, drafting and brainstorming

    Adapt prompting strategies based on task type (analysis, research, drafting, brainstorming)

    19 min read
    3.5% of exam
    7 sources
    Published 19 Oct 2026
    Docs as of 14 Oct 2026

    What you will be able to do

    • Pick the brief elements that matter most for an analysis, research, drafting or brainstorming request
    • Write an analysis prompt that lets Claude say the evidence is not enough instead of filling the gap
    • Give a research request, including one sent to Research mode, the scope, context and verification instructions it needs
    • Supply voice samples and real facts for a draft, and set up a draft, review, refine chain
    • Ask for breadth when brainstorming and run the evaluation as a separate step
    • Split a request that covers several task types so each part gets its own prompt

    Key concept

    Task-shaped prompt — Analysis, research, drafting and brainstorming fail in different ways, so each needs different context. A task-shaped prompt puts most of its detail into the elements that prevent that task's typical failure: invented conclusions, unfocused searching, generic prose or a narrow set of ideas.

    1.Why each task type needs a differently shaped prompt

    Four common requests look alike but go wrong in different ways. Ask Claude to analyse a quarterly report, research a market, draft a grant proposal or brainstorm campaign ideas, and the main risk changes each time. An analysis can reach conclusions the data does not support. Research can wander across topics nobody needed or treat one source as settled. A draft can come back in a polite, generic voice with invented details. A brainstorm can stop at five safe ideas. A prompt for one of these tasks puts its effort into blocking that particular failure.

    The general advice still holds for every type: an explicit brief: task, audience, format, constraints, depth, written so that a colleague with little context could follow it. Task type changes which part of the brief carries the weight. Analysis needs the evidence and a focus. Research needs a scope. Drafting needs a voice. Brainstorming needs range. Giving a reason for an instruction helps with all of them, because Claude generalises from the explanation instead of applying the rule mechanically.

    Where the weight of the brief goes for each task type
    Task typeBrief element that matters mostTypical failure when it is missingInstruction that prevents it
    AnalysisThe evidence itself, plus the criteria to judge it byConclusions the data does not supportPaste the data, name the focus areas, allow "there isn't enough information"
    ResearchScope: who you are, what you need, time frame, geographyBroad, generic or single-source answersDefine success criteria and ask for verification across several sources
    DraftingVoice and the actual factsGeneric tone and invented detailsSupply past work as voice samples and the real facts; review claims against them
    BrainstormingRangeA few obvious ideasAsk explicitly for breadth, then evaluate in a separate step

    Checkpoint 1 of 9· Match them up

    Each task type tends to fail in its own way. Match each task type to the prompt instruction that best prevents its typical failure.

    Tap a term, then the definition that fits it.

    Sources1

    2.Analysis: provide the evidence, name the focus, allow "not enough data"

    An analysis prompt starts with the material being analysed. Claude can only judge a report, a dataset or a policy that is actually in the prompt, so paste it or attach it, and wrap it in tags such as <report> so it stays separate from your instructions. For long inputs, put the documents at the top and the question below them. Then name the criteria. "Analyse this report" leaves Claude to guess what matters. "Focus on financial projections, integration risks, and regulatory hurdles" tells it what to look at.

    The instruction analysis prompts most often leave out is an explicit way out. Grounding a prompt in pasted source material and defining scope tells Claude what to work from. Analysis needs one more thing: permission to say the material is not enough. Without it, a question the data cannot answer still gets an answer, because filling the gap is easier than refusing. Write the escape route into the prompt, for example: "If the data is insufficient to draw conclusions, say so rather than speculating." Then a missing marketing-spend column shows up as a stated limitation, not a made-up cause.

    An analysis prompt with the evidence in tags, named focus areas, and permission to say the report is not enoughtext
    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."

    Two more instructions keep a long analysis tied to its evidence. First, ask Claude to extract the relevant quotes before it analyses, and to base its analysis only on those quotes. This makes it attend to what the document actually says and ignore the rest. Second, tell it to use only the provided documents and not its general knowledge, so background assumptions do not slip into a conclusion about your data. Both also make the output easier to audit: every claim points back to a passage you can check. A balanced analysis is built the same way. Ask for the strongest supporting evidence and the weakest counterpoints, each tied to a quote.

    Checkpoint 2 of 9· Check yourself

    A colleague's analysis prompt pastes a vendor report at the top, wrapped in report tags, then asks Claude to focus on pricing, contract terms and support coverage. The output confidently describes the vendor's support coverage, but the report never mentions support. Which change to the prompt most directly addresses this?

    Sources213

    3.Research: scope it, then make Claude check and weigh what it finds

    A research prompt is mostly scope. A broad question gets a broad answer. A nonprofit director asking "what housing policies matter in Seattle?" learns far less than one who explains who the organisation serves, why the move from Portland matters to its mission, and what they need to know: impact on beneficiaries, funding implications, compliance requirements. Add a time frame and a geographic boundary, and say what counts as a successful answer. Claude then knows when it has found enough.

    Reliability needs to be asked for, not assumed. A factual claim that several independent sources could confirm should be checked against more than one before anyone treats it as settled, and the prompt should say so: ask Claude to verify information across multiple sources and to cite them. The quick alternatives work against you. Answering from the first source found, stating a confident answer however many sources were checked, or dropping the source list to keep the reply short all remove the evidence a reader needs to judge the answer. Even after cross-checking, flag the specific claims you will verify yourself, look for gaps, and ask how recent the information is.

    Open-ended questions need a structured way to search. When the answer is uncertain, such as why a competitor is gaining share, tell Claude to develop several competing hypotheses as it gathers data, track its confidence in each in its notes, and regularly critique its own approach. That keeps rival explanations alive until the evidence separates them. Committing to one theory early turns later searches into confirmation. Running one broad search and answering from it skips the weighing altogether. Throwing away intermediate findings loses the record that shows why one explanation won.

    Research mode carries out this kind of task for you. It is built for comprehensive information gathering that takes five or more tool calls over one to three minutes, and it returns a longer, cited report. What it cannot provide is your context. The prompt you give Research still has to say who you are, what decision the report supports, what period and region count, and what a good answer contains. Otherwise it collects information competently on a question nobody asked.

    Checkpoint 3 of 9· Check yourself

    Planning a move from Portland to Seattle, a housing nonprofit's director turns on Research mode and types: "What housing policies matter in Seattle?" The report that comes back is long and cited, but too general to act on. Which change to the prompt would help most?

    Checkpoint 4 of 9· Exam question

    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?

    Checkpoint 5 of 9· Exam question

    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?

    Sources415

    4.Drafting: supply your voice and your facts, then review in stages

    A drafting prompt has to supply the two things Claude cannot know: how you sound and what is actually true about you. Left to itself, a draft falls back on a generic, competent tone and fills gaps with plausible details. Voice is easier to show than to describe, and examples are one of the most reliable ways to steer tone and style. Upload past work that succeeded, such as last year's funded proposal or a newsletter readers responded to, and ask Claude to match it.

    Facts come from you as well. Give Claude your track record, partnerships, staff expertise and real community relationships, along with the funder's priorities and hard requirements such as word count and required sections. A grant proposal written without them may read well and still be unusable, because the outcomes it cites are not yours. When you revise, be specific: "make this more compelling" gives Claude nothing to act on, while "add these specific details about our community relationships" does. Check the result for invented data points and for language that sounds generic instead of like your organisation.

    A drafting prompt restricted to supplied facts, with a self-check that removes any claim it cannot supporttext
    Draft a press release for our new cybersecurity product, AcmeSecurity Pro, using only information from these product briefs and market reports.
    <documents>
    {{DOCUMENTS}}
    </documents>
    
    After drafting, review each claim in your press release. For each claim, find a direct quote from the documents that supports it. If you can't find a supporting quote for a claim, remove that claim from the press release and mark where it was removed with empty [] brackets.

    When a draft has to pass a review, such as a style-guide check, make each stage its own call. The first prompt writes the draft. The second reviews that draft against the style guide and gives feedback. The third revises the draft using that feedback. Each prompt does one job, and its output feeds the next, so you can inspect and log the review on its own. Asking for draft, review and revision in one response hides the review inside a single answer. Skipping the review gives a polished draft that nobody checked against the guide. Generating ten versions and choosing one replaces the check with a guess. Chaining adds latency, and that is the cost of making each step easier to do accurately.

    Checkpoint 6 of 9· Fill the gap

    This drafting prompt is meant to stop Claude from inventing details in a press release. Which word fills the blank so that the draft stays tied to the supplied facts?

    Draft a press release for our new cybersecurity product, AcmeSecurity Pro, using  ?  information from these product briefs and market reports.
    <documents>
    {{DOCUMENTS}}
    </documents>
    
    After drafting, review each claim in your press release. For each claim, find a direct quote from the documents that supports it. If you can't find a supporting quote for a claim, remove that claim from the press release and mark where it was removed with empty [] brackets.

    Checkpoint 7 of 9· Exam question

    A content team wants Claude to draft a press release, have the draft checked against their style guide, and then produce a revised version based on that check, with each stage logged separately for review. How should this drafting workflow be set up?

    Sources613

    5.Brainstorming: ask for range, then judge in a separate step

    Brainstorming is the one task type where the prompt should ask for more than you will use. Claude tends to do what is asked, so to get breadth you have to request it explicitly. The same principle shows up in a coding example: "create an analytics dashboard" gets a basic result, while adding "include as many relevant features and interactions as possible" gets an ambitious one. For ideas, ask for many options, ask for options that differ from each other, and name the angles to cover, such as cheap versus expensive, short-term versus long-term, or customer-facing versus internal.

    Constraints still go in during the divergent phase. The difference is that they set the boundaries without starting to rank ideas. A startup with a $500 marketing budget gets more usable ideas by stating that budget, the audience and the timeline up front than by brainstorming freely and discarding most of the results. Reframing can produce range too. To generate safety features for a vehicle, first ask for the ways it could fail or cause an accident, then ask for the inverse of each one as a safety measure.

    Evaluation belongs in its own step. If you ask Claude to generate options and judge them in the same breath, ideas get screened out as they come up, and only the defensible ones make the list. Generate first. Then send the full list back with your criteria (cost, feasibility, time to implement, risk) and ask for a critical assessment of each, followed by a recommendation. This is the generator-and-evaluator pattern: one call produces, another evaluates and gives feedback. It works best when the evaluation criteria are clear, which is exactly what the second step adds.

    Checkpoint 8 of 9· Check yourself

    A marketing lead asks Claude, "Give me ideas for our product launch campaign," and gets back five safe, predictable suggestions. What should the next prompt change to get more range?

    Sources17

    6.Requests that span several task types

    Many real requests combine task types: research a funding landscape and then draft the proposal, or analyse churn data and then brainstorm retention ideas. One prompt covering both has to serve two purposes. The research half needs scope and verification. The drafting half needs voice and facts. Usually one half gets shortchanged. Applying deciding when to decompose a request here means splitting the request where the task type changes, so each part gets the prompt suited to it.

    The seam is also where you review. Between research and drafting, check the findings: confirm the claims you flagged and drop what you could not verify. Only then hand the confirmed facts to the drafting prompt, along with your voice samples. A drafting prompt that starts from checked research does not have to invent details, and a research prompt that does not also have to write polished prose can stay focused on scope and sources. The extra round trip costs time. In return, each step is easier for Claude to do accurately and easier for you to check.

    Checkpoint 9 of 9· Check yourself

    An executive director sends Claude a single request: "Research which foundations fund youth literacy programmes in our region, then write our grant proposal." The proposal comes back polished but cites funder priorities nobody checked and sounds nothing like the organisation. What is the best way to restructure the work?

    Sources34

    Sources

    (7)

    Every claim above is drawn from one of these pages, quoted as it was written on the date shown.

    1. 1.
      “Show your prompt to a colleague with minimal context on the task and ask them to follow it.”
      ↩︎ Why each task type needs a differently shaped prompt
      “Claude is smart enough to generalize from the explanation.”
      ↩︎ Why each task type needs a differently shaped prompt
      “ask Claude to quote relevant parts of the documents first before carrying out its task.”
      ↩︎ Analysis: provide the evidence, name the focus, allow "not enough data"
      “Provide clear success criteria: Define what constitutes a successful answer to your research question.”
      ↩︎ Research: scope it, then make Claude check and weigh what it finds
      “Encourage source verification: Ask Claude to verify information across multiple sources.”
      ↩︎ Research: scope it, then make Claude check and weigh what it finds
      “As you gather data, develop several competing hypotheses.”
      ↩︎ Research: scope it, then make Claude check and weigh what it finds
      “Examples are one of the most reliable ways to steer Claude's output format, tone, and structure.”
      ↩︎ Drafting: supply your voice and your facts, then review in stages
      “Include as many relevant features and interactions as possible. Go beyond the basics to create a fully-featured implementation.”
      ↩︎ Brainstorming: ask for range, then judge in a separate step
      “Be specific about the desired output format and constraints.”
      ↩︎ Brainstorming: ask for range, then judge in a separate step
      “As you gather data, develop several competing hypotheses.”
      ↩︎ Exam trap 2
      “Create an analytics dashboard. Include as many relevant features and interactions as possible.”
      ↩︎ Checkpoint
    2. 2.
      “Explicitly give Claude permission to admit uncertainty. This simple technique can drastically reduce false information.”
      ↩︎ Analysis: provide the evidence, name the focus, allow "not enough data"
      “Explicitly instruct Claude to only use information from provided documents and not its general knowledge.”
      ↩︎ Analysis: provide the evidence, name the focus, allow "not enough data"
    3. 3.
      “If the data is insufficient to draw conclusions, say so rather than speculating.”
      ↩︎ Analysis: provide the evidence, name the focus, allow "not enough data"
      “Each prompt handles one stage, and the output feeds into the next instruction.”
      ↩︎ Drafting: supply your voice and your facts, then review in stages
      “This approach trades latency for higher accuracy by making each individual task easier.”
      ↩︎ Requests that span several task types
      “If the data is insufficient to draw conclusions, say so rather than speculating.”
      ↩︎ Exam trap 1
      “This approach trades latency for higher accuracy by making each individual task easier.”
      ↩︎ Exam trap 4
      “Each prompt handles one stage, and the output feeds into the next instruction.”
      ↩︎ Prediction
      “Chaining breaks down complex tasks into smaller sequential steps with separate prompts. Each prompt handles one stage, and the output feeds into the next instruction.”
      ↩︎ Checkpoint
    4. 4.
      “Flag specific claims for verification, notice gaps, and question recency of information”
      ↩︎ Research: scope it, then make Claude check and weigh what it finds
      “AI accelerates research but doesn't replace expertise”
      ↩︎ Requests that span several task types
      “Don't just ask broad questions—explain who you are, who you serve, and what you specifically need to know.”
      ↩︎ Key concept
      “Don't just ask broad questions—explain who you are, who you serve, and what you specifically need to know.”
      ↩︎ Checkpoint
    5. 5.
      “Research is optimal for comprehensive information gathering requiring five or more tool calls over 1-3 minutes”
      ↩︎ Research: scope it, then make Claude check and weigh what it finds
      “The research feature can synthesize information across multiple sources with proper citations and produces a longer final report.”
      ↩︎ Research: scope it, then make Claude check and weigh what it finds
    6. 6.
      “Upload past successful work to help AI understand your organization's voice and approach—this saves time and improves authenticity”
      ↩︎ Drafting: supply your voice and your facts, then review in stages
      “Inject the details AI can't know: your track record, partnerships, staff expertise, and actual community relationships”
      ↩︎ Drafting: supply your voice and your facts, then review in stages
      “Instead of "make this more compelling," say "add these specific details about our community relationships"”
      ↩︎ Drafting: supply your voice and your facts, then review in stages
      “Check if the language reflects your organization's actual voice or sounds generic”
      ↩︎ Drafting: supply your voice and your facts, then review in stages
      “Upload past successful work to help AI understand your organization's voice and approach—this saves time and improves authenticity”
      ↩︎ Exam trap 3
      “Upload past successful work to help AI understand your organization's voice and approach—this saves time and improves authenticity”
      ↩︎ Checkpoint
    7. 7.
      “one LLM call generates a response while another provides evaluation and feedback in a loop.”
      ↩︎ Brainstorming: ask for range, then judge in a separate step

    Spotted a mistake, or was something unclear? Tell us.