Home › Study guides › CCAO-F › Domain 1 › Lesson 1.4

CCAO-F · Domain 1 · 14% of the exam · Lesson 1.4 · 19 min read

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

Why analysis, research, drafting and brainstorming each need a different prompt and a different test, and how to run them in order on one real project.

Written against objective 1.4 of the official CCAO-F exam guide (Version 1.0, effective July 2026). An independent resource, not affiliated with Anthropic; the practice questions are written from scratch.

1.4.1 Why one kind of prompt fails four different jobs

Your company's annual engagement survey closed on Friday. You're the HR business partner at a manufacturer with about 1,100 people across two plants and three shifts. You now owe four things: results by department for leadership, a view on how other employers keep shift workers, a shortlist of initiatives for next year, and an announcement to all staff.

You paste the department results into Claude (the export holds no names) and ask: "What's going on with the night shift, and what should we do about it?" The reply is fluent and well organised. It says engagement "fell sharply in Assembly", quotes an industry figure on night-shift turnover, suggests five initiatives, and ends with a paragraph ready for the announcement.

Look closer and each part fails its own test. The Assembly claim shows no working, so you can't tell which measure it used or whether it noticed that only a third of the night shift replied. The industry figure has no source or date. The initiatives read like recommendations nobody has checked. And the announcement would repeat all of it to 1,100 people.

None of this is a wording problem. You asked for four kinds of work at once, and each has a different goal. Analysis finds what your own data shows; research finds what is known elsewhere; brainstorming widens the options; drafting puts settled content into the right words for a reader. That goal, the task type, decides what the prompt must supply and how you judge the answer.

1.4.2 Name the task type before you write the prompt

Here is the question to ask before you type anything: what is this request actually FOR? The four types are easy to blur, because Claude answers all of them in the same fluent, confident prose, whether a sentence came from your data, a checked source or its own guesswork.

Think of four colleagues on the survey project: a data analyst, a librarian, a workshop facilitator and a copywriter. You'd brief each one differently and judge each by a different standard; you'd never take the facilitator's flip chart to the board as audited figures. Claude can do all four jobs, but it can't tell which one you want unless the prompt says so.

Memorise the goal of each type and the last column, the one people skip; the middle column follows from them. The six parts of a good brief (task, context, audience, material, constraints and output) apply to every type. The task type decides which of them carry the weight.

1.4.3 Analysis: ask for the working, not just the answer

Start with the leadership request: which departments and shifts have the lowest engagement, and how has that changed? It hides several choices: what counts as engaged, whether to compare averages or percentages, what to do with a team where eight people replied. Claude only knows what is in front of it. If the prompt doesn't settle those choices, Claude settles them silently, and the conclusion reads just as confidently either way.

So an analysis prompt gives Claude the data, what each field means, how the measure is defined and the decision the answer feeds. Then it asks for the working: the calculations, the assumptions and the rows behind each finding. Anthropic's prompting advice also suggests giving Claude permission to say the data is insufficient rather than speculate. In the prompt below, look at the third paragraph, which asks for the working first, and the last, which keeps findings and explanations apart.

Goal: find which departments and shifts have the lowest engagement this year, and how that has changed since last year. Leadership will use this to decide where to focus next year.

Data: the attached export, one row per department and shift, with no names. Columns: Dept, Shift, Headcount, Responses, and % favourable for questions Q1 to Q12, this year and last year. "Favourable" means a 4 or 5 on the 5-point scale. Engagement is the average of Q1 to Q12.

Work through the calculations step by step before you give any findings. Show each group's engagement score, its response rate and its change from last year.

For each finding, name the rows and figures it rests on, and list any assumption you made. Do not rank groups with fewer than 10 responses; say they are too small to compare.

Give two separate lists: FINDINGS (what the numbers show) and POSSIBLE EXPLANATIONS (ideas the data cannot prove). If the data cannot answer part of the question, say so rather than guessing.

Why two lists? "Night-shift Assembly fell 14 points" can be checked against the export. "Probably the new rota" cannot; it's a hypothesis for a focus group to test. Written in one paragraph they look equally solid, and leadership will repeat whichever sounds most interesting.

Analysis also benefits from room to think. Claude's thinking setting (called extended thinking on some Help Center pages) lets it break a problem down and plan before it answers, and a higher effort level makes its responses more thorough. The Help Center lists detailed document analysis among the tasks these settings help most. Where thinking isn't already on, turn it on, or ask Claude to work through the problem step by step.

1.4.4 Research: scope the question, then follow the sources

The second request, from the operations director, is a different kind of task: what do other employers do to keep shift workers? The answer isn't in your export. With no sources and no search, Claude answers from its training, which may be out of date. The Help Center also warns that it can produce quotes that look authoritative but aren't grounded in fact. The turnover figure in the first reply might be real, half-remembered or invented, and it would read the same.

This is the librarian's job: you want the report with the page marked, not "I read somewhere that...". So a research prompt scopes the question: which workers, which sectors and countries, how recent, and what counts as evidence that a practice worked. It says where the evidence should come from: files you supply, such as uploaded or connected documents, or a web search or the Research feature. And it asks for a source and date for every claim, keeping what a source actually reports apart from what Claude concludes from it.

Anthropic's prompting guide also suggests asking Claude to check claims across more than one source. Dates matter because shift patterns, pay rules and labour markets change; a practice that worked for retail staff ten years ago says little about your night shift. Finally, open the key sources before a figure reaches the leadership pack. The original often holds context the summary left out, and a confident summary is not evidence.

1.4.5 Brainstorming widens, drafting narrows

The last two types pull in opposite directions, which is why mixing them up hurts. Brainstorming is for when you don't yet know the options. Leadership wants initiatives for next year, and the worst outcome is a shortlist of the three ideas everyone already had.

So a brainstorming prompt asks for quantity and variety: twenty ideas across scheduling, pay, recognition, development and the working environment, including a few that look impractical. It relaxes the usual constraints, because a budget filter at this stage removes ideas before anyone has seen them. And the output is labelled honestly: ideas, not findings. If an idea arrives with a claim attached ("proven to cut turnover by 30%"), that claim is unchecked research, not a fact.

Then you converge: you narrow the list, and the constraints come back as explicit criteria. Ask Claude to score the ideas against criteria you state: cost, how soon you could start, whether it reaches all three shifts, whether it answers something the analysis found. "Pick the best five" leaves Claude to invent its own criteria.

Drafting is the reverse: every constraint is tight from the start, built around its readers. Many staff will see the announcement on a noticeboard or hear it at a shift briefing rather than open an email. So the prompt names the audience, the purpose (share the results honestly and say what happens next), the channel, tone, length and format. It includes one sample of the voice, such as last year's well-received update, marked "copy the tone, not the content". And it lists the facts to use, only these: checked figures and decisions leadership has actually made.

1.4.6 The wrong strategy, and the right order

Now the classic failure: one task type's prompt used for a different kind of question. A shift supervisor forwards a question from her team: "Can our managers see how we answered?" You ask Claude to "write a short, reassuring FAQ answer about survey anonymity". Back comes a warm paragraph promising that no one at the company can ever see individual answers. It's fluent, kind and confident. It may also be false: perhaps HR administrators can filter results by shift and department, which in a small team comes close to naming people.

You wrote a drafting prompt, so you got drafting: the most reassuring plausible words. But this question has a right answer, set by the survey provider's privacy notice and your data-protection policy. A question whose answer an approved source or its owner decides is an authoritative question, and it calls for analysis of that source, not drafting. In the prompt below, notice three moves: the sources are fixed, every point needs a quote, and gaps go to a named person instead of being filled.

Question from staff: "Can our managers see how we answered the engagement survey?"

This is a policy question, not a writing task. Answer it only from the two attached documents: the survey provider's privacy notice and our employee data-protection policy.

For each point in your answer, quote the sentence it rests on and name the document.

If the documents do not settle a point, write "Not covered" and list it as a question for our data protection lead. Do not fill gaps with what is usual for surveys.

Only after that, write a three-sentence version for staff that says nothing beyond the quoted text.

Most real work mixes types, so it needs a sequence, not one request that asks for everything at once. First define the objective and the trust boundaries: which decisions stay with leadership, and which questions only the data protection lead can answer. Then analyse, brainstorm options with research feeding in what has worked elsewhere, validate, and only then draft. Each step is a different TYPE of task, and gets the prompt that fits it.

The order matters because each stage inherits the errors of the one before. A brainstorm built on an unchecked analysis chases the wrong department, and a draft written before validation sends mistakes to 1,100 people. Validation is where you recheck the numbers, open the sources and test the shortlist with the plant managers.

1.4.7 The exam traps

Every trap here is a mismatch between the prompt and the task's real goal.

• ✗ Answering a policy question with a drafting prompt ("write a reassuring answer"). ✓ Treat it as analysis of the approved source: answer only from it, quote the clause, and route what it doesn't cover to the owner.
• ✗ Asking for conclusions from data with no definitions and no working. ✓ Supply fields and definitions; ask for calculations, assumptions and evidence, with findings apart from hypotheses.
• ✗ Accepting a research answer because it sounds well informed. ✓ Scope the question, supply sources or let Claude search, ask for a source and date per claim, and check the key ones.
• ✗ Presenting brainstormed ideas, or the claims attached to them, as findings. ✓ Label them as ideas, narrow them against stated criteria and validate before anyone relies on them.
• ✗ Asking for everything in one request, or drafting before the facts are checked. ✓ Sequence the stages, each with its own prompt, and draft last, from verified facts, for a named reader and channel.

1.4.8 Put it together: run the survey project one task type at a time

You now have every piece: name the task type, brief for its goal, and when a job mixes types, run them in order and validate before you draft. The quickest way to feel the difference is to watch a strong analysis prompt work, then take its ingredients away.

Domain 2 turns the last column of the task-type table into a discipline: fact-checking techniques (2.3) and deciding when a human must review (2.4). Domain 3 covers choosing between web search, thinking and the Research feature (3.1). And in Domain 5, task-specific prompt patterns can live in a Project's instructions (5.1), so a whole team starts from the right one.

Key takeaways

• ✓ Name the task type first: analysis, research, brainstorming and drafting differ in goal, so each needs different material in the prompt and a different test for the result.
• ✓ For analysis, supply the data, definitions and question, and ask for calculations, assumptions and evidence, with findings kept apart from hypotheses.
• ✓ For research, scope the question, say where evidence should come from, ask for a source and date per claim, and verify the key claims before use.
• ✓ For brainstorming, ask for many varied ideas with relaxed limits, label them as ideas, and converge on criteria you state; a brainstormed claim is never a fact.
• ✓ For drafting, name the audience, purpose, channel, tone, length and format, show one sample of the voice, and give only approved facts.
• ✓ A drafting or brainstorming prompt for an authoritative question, such as a policy answer, yields a fluent, confident and unsupported reply; ground it in the approved source instead.
• ✓ Real work mixes types: define, analyse, brainstorm, validate and then draft, with a check before anything reaches the draft.

Check your understanding

4 questions written for this lesson, then one from the CCAO-F question bank on the same topic. Every answer option is explained, including the ones you did not pick. Nothing is stored.

Sources

• Prompting best practices
• Prompt engineering best practices
• When should I use web search, extended thinking, and research?
• Change the model, effort, and thinking settings
• Use research on Claude
• Claude is providing incorrect or misleading responses
• The CCAO-F exam guide, summarised