By Amotion AIPrompting and Task Execution: CCAO-F domain 1 study guide
CCAO-F · Prompting and Task Execution (14% of the exam)
Prompting and Task Execution is domain 1 of the Claude Certified Associate Foundations exam, 14% of the scored items. It tests whether you can turn a work request into a prompt that gets a usable result, split a large job into steps you can check, and fix a weak result by changing the part of the prompt that caused it.
What the official guide covers
The Claude Certified Associate Foundations exam guide (version 1.0, effective July 2026) lists four tasks under domain 1, Prompting and Task Execution:
| What the guide lists | What it means in practice |
|---|---|
| Create effective prompts for business and technical tasks | State the goal, the reader, the source material, the limits and the format you want back |
| Apply task decomposition to structure complex requests | Split a large job, such as a vendor comparison or a board pack, into steps you check one at a time |
| Iterate prompts to improve output quality | Read the result, name what is wrong, and change the part of the prompt that caused it |
| Adapt prompting to the task type (analysis, research, drafting, brainstorming) | Ask for different things: method and evidence for analysis, sources for research, tone and length for drafting, range for brainstorming |
What a good work prompt contains
Anthropic's prompting guidance describes Claude as a brilliant new employee who does not yet know your norms or workflows. Its test is simple: show the prompt to a colleague with no background on the task. If they would be confused, Claude will be too.
A full work prompt has up to seven parts:
| Part | What to write | Example |
|---|---|---|
| Role | Who Claude should act as | "You are an operations analyst" |
| Goal | What the output is for | "A one-page update for the steering committee" |
| Reason and reader | Who reads it and what they will decide | "They decide on Thursday whether to extend the pilot" |
| Source material | The notes, files or data to use, kept apart from your instructions | Notes pasted inside <notes> tags, or an attached file |
| Limits | Length, what to leave out, what not to invent | "Use only figures in the notes. Write [missing] for any gap" |
| Format | The shape of the answer | "Three headings: Progress, Risks, Decision needed" |
| Example | A sample of the tone or layout you want | Last month's approved update |
Not every prompt needs all seven. A quick question needs a clear task and perhaps one limit; a client deliverable needs every part. Context is the part people leave out most: Claude cannot see your inbox or your last meeting unless you paste it or connect it.
Anthropic's AI Fluency framework calls this Description and splits it three ways: the product (what you want back), the process (how Claude should work, such as "list the criteria before scoring") and the performance (how Claude should behave with you, such as "challenge my assumptions" or "keep replies short").
Give the reason behind a rule, because Claude generalises from it: saying a reply will be read aloud explains why ellipses must go. For long documents, put the document first and the question last, and ask for supporting quotes before the answer.
Before and after: a weak prompt rewritten
Before:
Summarise these notes for the team.
After:
I am preparing for Thursday's steering committee. They will decide whether
to extend our customer onboarding pilot for another quarter.
Using only the meeting notes inside <notes>, write a one-page update with
three headings: Progress, Risks, Decision needed.
- Readers: senior managers who were not at the meetings. Plain language,
no internal project codes.
- Keep it under 300 words.
- Use only figures that appear in the notes. If a figure the committee
would expect is missing, write [missing] instead of estimating it.
- Under Decision needed, list the options and the staff time each one
needs, if the notes say.
<notes>
[paste notes here]
</notes>
What changed: purpose and reader stated, source limited to the notes, format fixed, and a rule for missing data.
How to break a large request into steps
One prompt that asks for everything at once gives you one large answer and no way to see where an error came in. Anthropic's guidance says splitting work into a chain of prompts is still useful when you need to check intermediate results. A vendor comparison might run like this:
- Extract. "From the three proposals, list each vendor's pricing model, contract length and support terms as a table. Leave a cell blank if the proposal does not say."
- Check. You compare the table with the proposals and fix any errors.
- Analyse. "Using this table, score each vendor against our five requirements and explain each score in one line."
- Draft. "Write a half-page recommendation for the finance director, based on the scores."
Each step's output becomes the next step's input, and you check it before moving on. Within one step, you can also ask for a draft, then a review against your criteria, then a revision.
When several outputs share one source
A facilities manager must turn a new building access policy into a staff email, a reception desk script and a short note for directors. The tempting order is to draft all three at once. The safer order:
- Extract every change in the policy and what it means for a person arriving at the building.
- Check that list against the policy yourself.
- Draft the three outputs, each from the checked list, one at a time.
If the extraction misreads a rule, drafting first copies the same mistake into three documents. Build and check the shared foundation, then let the outputs branch from it.
One chat or several
Keep steps that depend on each other in one chat so each step sees the earlier results. Open a new chat when a step does not depend on the others, or when the chat has run long enough that early instructions are losing force; carry over a short checked summary rather than the whole history.
How to iterate: change the cause, not the wording
| What is wrong with the output | Likely cause in the prompt | Change to make |
|---|---|---|
| Generic, could apply to any company | No reader or context | Say who reads it and what they will do with it |
| Wrong length or shape | No format given | State the headings, length, and table or prose |
| Figures that are not in your source | No rule for missing data | Limit Claude to the source and tell it how to mark gaps |
| Tone is off | Nothing to copy | Paste an approved example and label it as the example |
| Half the request is missing | Too many asks in one prompt | Split the job into steps |
| Totals or counts that look wrong | Arithmetic written as prose | Ask Claude to calculate it with code execution |
| Wrong action, such as a summary when you wanted a comparison | Task verb unclear | Use one clear verb: compare, rank, draft, extract |
| Mostly right, one section weak | Nothing wrong with the rest | Ask for a revision of that section only |
Give specific feedback in the same chat ("Cut Risks to three bullets and keep every figure") rather than "make it better". Pressing regenerate repeatedly changes the wording, not the cause. Change one part per round: if you rewrite the whole prompt, you lose what worked and cannot tell which change helped. When a chat has gone through many corrections, start a new one with the improved prompt.
Know when to stop. When a round produces only small changes, finish the draft by hand. Aim for a result you can use, not a perfect prompt.
Matching the prompt to the task type
Each task type needs a different balance between control and freedom. Decide what you must fix and what you can leave to Claude.
| Task type | Ask for (tighten) | Leave open | Watch for |
|---|---|---|---|
| Analysis | The criteria, the standard to measure against, the data, assumptions kept apart from findings | Phrasing | Conclusions the data does not support |
| Research | The exact question, the scope, citations, and a flag on anything unverified | How Claude organises the synthesis | Claims with no source you can open |
| Drafting | Reader, tone, length, format and an example | Word choice | Confident wording that adds facts not in your notes |
| Brainstorming | The goal and one or two limits only | Number and direction of ideas | Judging ideas in the same step as generating them |
For research, one current fact needs only web search switched on in the chat. A question needing many sources suits Research, on paid plans, which runs searches that build on each other and returns citations (web search must be on). Those citations link to pages you can open; a reference Claude writes from training alone can look just as precise and still be wrong.
Rules that decide exam answers
- Specific beats emphatic. The right option adds the reader, purpose, source and format. Wrong options add "please be accurate", capital letters or a longer request with the same gaps.
- Calibrate in both directions. Missing detail gives generic output, but heavy limits on a brainstorm kill the range you wanted. For idea generation, the right answer loosens the limits and filters afterwards.
- Split the job when the steps need checking. If an error could come from any of several sub-tasks, the answer is to separate them and check each output.
- Fix the cause, not the symptom. Regenerating, switching tools or rewording politely are tempting distractors when the prompt is missing context or a format.
- Say what to do when information is missing. A rule such as "write [missing]" prevents invented figures better than "do not make things up".
- Check the shared step first. When several outputs come from one source, the right answer verifies the extraction before any drafting; drafting all outputs at once is the tempting wrong option.
Where it appears in the exam
Domain 1 carries 14% of the exam, so expect around 8 of the 60 items. Items are multiple choice or multiple response, and each one states how many answers to select. Questions describe a professional with a request and a weak result, and ask which prompt change, way of splitting the task or approach for the task type works best.
Two sample questions
These are original Timo practice questions. They are not official exam questions.
Build exercise
- Pick a task you do every week, such as a status update or a meeting summary. Run the one-line prompt you would normally use and save the result.
- Rewrite the prompt using the parts above that the task needs. Run it in a new chat and list every difference between the two outputs.
- Take a larger task and split it into three or four steps. Check the first output before running the second, and note any error you caught that a single prompt would have hidden.
Practise this topic
- Claude Certified Associate practice exam: free, 20 questions, no sign-up
- CCAO-F study guide: all topics
- Next topic: Output Evaluation and Validation
Sources
- Claude Certified Associate Foundations Exam Guide, version 1.0, effective July 2026 (Anthropic), domain 1: Prompting and Task Execution
- Anthropic documentation: Prompting best practices
- Claude Help Center: Use research on Claude
- Claude Help Center: Enabling and using web search
- Claude Help Center: Create and edit files with Claude