Prompting and Task Execution

What this domain tests

14% of the exam — roughly 8 of 60 items. The blueprint lists four tasks: create effective prompts for business and technical work, decompose complex requests, iterate to improve quality, and adapt strategy to the task type (analysis, research, drafting, brainstorming).

First-hand reports describe the Associate exam as judgment-heavy rather than recall-heavy: scenarios where several options are defensible and one is clearly best. In this domain that means workplace vignettes — a draft that missed the mark, a template that fails in other people's hands, a brainstorm that came back flat — where you pick the intervention a skilled practitioner would make. The recurring shape: the output is technically fine but unusable, and the right answer is the one that adds the missing specification, not the one that switches models, adds emphasis, or retries and hopes.

Think of prompting the way managers think of delegation. A prompt is a brief to a very capable new colleague who knows nothing about your company, your audience, or your standards. Every technique in this domain is a delegation technique wearing an AI costume.

Theory outline

1. Anatomy of an effective prompt

Anthropic's golden rule of clear prompting: show your prompt to a colleague with no context on the task. If they couldn't produce what you want from it, neither can Claude. The exam operationalizes this as component-spotting — which piece of the brief is missing?

Working rules the exam rewards:

• Be specific and direct. "Summarize this report" is a wish; "a 5-bullet summary of the three biggest risks, for non-financial executives, each bullet under 20 words" is a specification. Specificity is the cheapest quality lever that exists.
• Prefer positive instructions. "Write in short declarative sentences" beats "don't be verbose." Telling Claude what to do gives it a target; telling it what to avoid leaves the target undefined.
• Separate instructions from content. When a prompt mixes your instructions with pasted material, use clear delimiters — headings or XML-style tags like a "document" tag around the pasted text. Without separation, sentences inside the document can read as instructions, the classic failure of shared templates.
• Role prompting ("You are an experienced procurement manager") frames perspective, vocabulary, and judgment standards. It does not unlock hidden capability, change speed, or substitute for task specifics — three claims distractors love.
• Few-shot examples. Instructions describe; examples demonstrate. When the requirement is "same structure every time, across many runs or many users," an example of the exact desired output is the most reliable lock. One to three well-chosen examples that show the edge cases beat ten that show the same happy path.
• Edge-case rules. Robust prompts state the fallback: "if the owner is not named, list the item under Unassigned." The alternative is Claude's best guess, delivered in the same confident register as everything else.

Business vs technical tasks. Same anatomy, different emphasis. Business tasks live or die on context, audience, and stakes ("second outage this quarter, enterprise customer, offering 15% credit"). Technical tasks live or die on precise inputs, expected outputs, and edge-case handling ("input rows may have empty cells; output valid CSV with these exact headers; flag rows you could not parse instead of guessing"). The exam expects you to know which levers matter for which kind of work.

2. Task decomposition

The core insight: a prompt asking for five things gives each thing a fifth of the attention, and an error in step one silently poisons steps two through five. Decomposition — splitting complex work into sequenced, reviewable steps — is how practitioners get reliability that mega-prompts cannot deliver.

Signals a task needs decomposition:

The recurring patterns, each with a name worth recognizing in options:

• Sequential chaining. Step 1's output becomes step 2's input, with human review between steps. Summarize → extract → draft. Review at each seam catches errors while they are still cheap.
• Outline-first (checkpoint early). For any large deliverable, have Claude produce the outline or plan, approve it, then generate the full version. Fixing a wrong outline costs one message; fixing a wrong 20-page draft costs an afternoon. Place human checkpoints where mistakes are cheapest to correct.
• Parallel fan-out, then synthesis. When many inputs need identical treatment (ten transcripts, six competitors), process each through the same extraction prompt into a fixed template, review the extractions, then run one synthesis step across the structured intermediates. The fixed intermediate format is what makes the synthesis reliable — Claude compares like with like.
• Extract-then-reason. Separate "pull the facts out of the source" from "reason about the facts." Verification becomes possible at the seam: you can check the extraction against the document before any reasoning is built on it.

What decomposition is not: splitting a request into fragments to evade safety review (a misuse pattern, not a technique), or splitting so finely that you spend more time orchestrating than the task is worth. Decompose along the seams where review adds value.

3. Iterating prompts to improve output quality

Claude keeps conversational context, so the default iteration loop is in-chat with directional feedback: what to keep, what to change, in which direction. "Keep the structure; make the second section half as long and lead with the cost figure" gives Claude a vector. "Make it better" produces arbitrary change — Claude has to guess which axis you meant.

The disciplines the exam tests:

• One variable at a time. When improving a prompt that will be reused, change one element per test. Change three things at once and you cannot attribute the improvement — or the regression.
• Test against representative inputs, not the tuning sample. A prompt polished all afternoon against one sample document tends to overfit it — instructions quietly reference that document's quirks. Before declaring a reusable prompt done, run it against a small saved set of varied, realistic inputs. This is the everyday, no-code version of evaluation practice: define what success looks like, keep test cases, check changes against them.
• The reset move. After many rounds of partially contradictory steering, in-chat iteration degrades — the accumulated corrections pull in different directions and outputs get worse, not better. Consolidate everything learned into one clean prompt and start a fresh conversation. Knowing when to stop iterating in place is the skill.
• Escalation ladder for a struggling prompt: clarify the instruction → add or improve an example → decompose the task → only then consider a stronger model. Options that jump straight to switching models when the brief is the problem are distractors nearly every time.
• Codify what works. A prompt that survives testing becomes a template: saved, named, shared, with a note on what inputs it expects. Team prompt quality compounds through libraries, not memory.

4. Adapting strategy to the task type

The blueprint names four task types. Each defines "good output" differently, so each rewards different prompting moves — and the exam loves scenarios where someone applies the wrong type's strategy.

Two type-specific mechanics worth internalizing:

• Reasoning before answering. For analysis and multi-factor decisions, instruct Claude to walk through the trade-offs before stating the recommendation. Reasoning-then-answer measurably beats answer-first on complex judgment; asking for the answer first and the reasons after gets you rationalization, not reasoning.
• Divergence vs convergence. Brainstorming is a divergence task: the prompt should widen the space (quantity targets, category quotas, "include two options you think are impractical"). Evaluation is a convergence task and belongs in a separate step with explicit criteria. Mixing them — "brainstorm ideas and tell me the best one" — collapses the divergence and produces five safe, similar options. And because models pick up on framing, mentioning the idea you already like anchors the whole session around it.

Worked problems

Problem 1. A sales director prompts: "Write a proposal for the Meridian account." Claude produces a polished proposal that could have been written for any client. What is the highest-leverage fix?

A. Regenerate until one version happens to fit B. Add the account specifics: Meridian's stated priorities, the deal history, the two objections raised on the last call, and the required proposal sections C. Switch to the most capable model available D. Append "make it persuasive and specific to this client" to the prompt

Approach: Diagnose which prompt component is missing. The output is generic — the context dial. Claude cannot know unstated business facts; the fix is supplying them. Why the others fail: A re-rolls the dice on a prompt that can't produce the target — the information isn't there to use. C misdiagnoses: no model tier can recover facts absent from the prompt. D adds an adjective, not information — "be specific" without specifics is a wish. Answer: B

Problem 2. An analyst must turn six competitor websites and two analyst reports into a competitive-landscape deck for leadership. Which structure produces the most reliable result?

A. One prompt: paste everything, "produce the deck" B. Extract each competitor into the same fixed template, review the extractions, then run a synthesis prompt across the structured summaries, then draft the deck section by section C. Draft the deck first and backfill facts afterward D. Give Claude all sources and let it choose its own structure with no interim review

Approach: Eight sources, one large structured output, cross-source comparison — three decomposition signals at once. The pattern is parallel fan-out with a fixed intermediate format, review at the seam, then synthesis. The fixed template is load-bearing: it makes the extractions comparable and checkable before anything is built on them. Why the others fail: A gives each source a fraction of the attention and any misreading propagates invisibly into the deck. C inverts the dependency — conclusions first, evidence later is rationalization by design. D removes the checkpoints, which is where the reliability came from. Answer: B

Problem 3. A contracts manager spends an afternoon refining a clause-flagging prompt against one sample contract until it works perfectly. On the next three real contracts, it misses obvious clauses. What happened, and what's the fix?

A. Model variance — regenerate and results will improve B. The prompt overfit the tuning document; re-test against a representative set of contracts and rewrite the instructions that quietly depended on the sample's structure C. The task is too hard for the model — escalate tiers D. Add "be thorough and careful" to strengthen the prompt

Approach: Perfect on the tuning sample, poor on everything else is the signature of overfitting — instructions that referenced the sample's section names, ordering, or phrasing. The remedy is the testing discipline: a small saved set of varied real inputs, and generalized instructions that survive all of them. Why the others fail: A misreads a systematic failure as noise — the misses will recur. C escalates before diagnosing; the prompt is the defect. D adds emphasis, not specification — exhortation never fixed a prompt. Answer: B

Problem 4. A marketing manager prompts: "Brainstorm launch campaign ideas — I'm thinking something like a referral program." Claude returns five referral-program variations. What should the manager change?

A. Ask for ten more ideas in the same chat B. Re-ask without revealing the favored direction; request 20 ideas across named categories (partnerships, events, content, product-led, out-of-home) and defer evaluation to a second pass C. Switch to a model marketed as more creative D. Add "think outside the box" to the original prompt

Approach: Brainstorming is a divergence task and the prompt anchored it: mentioning the referral idea told Claude what kind of answer was wanted. The type-correct strategy removes the anchor, forces variety structurally (quantity + category quotas), and separates generation from judgment. Why the others fail: A extends the anchored session — more referral variants. C changes the tool when the prompt set the trap. D is emphasis without structure; the anchor is still in the prompt. Answer: B

Problem 5. An associate must brief leadership on a competitor that pivoted twice in the past year. Which prompting strategy fits this research task?

A. Ask Claude to answer from its knowledge, then polish the wording B. Have Claude use web search, require a citation for every factual claim, and instruct it to separate sourced facts from its own inference C. Add "make sure everything is current" to the prompt D. Ask the same question in three chats and keep the overlapping claims

Approach: Fast-moving subject → training knowledge is stale by construction, so the strategy is grounding: search for current sources, cite them so every claim is inspectable, and fence off inference from fact so the reader knows which is which. Why the others fail: A polishes potentially outdated content — fluency isn't currency. C is an instruction the model cannot obey; phrasing adds no knowledge. D confuses agreement with accuracy — three samples of the same stale training data will happily overlap. Answer: B

Problem 6. A finance lead asks Claude to compare two vendor proposals: "Which proposal is better?" The answer is confident but the CFO will ask "on what basis?" What prompt produces a defensible recommendation?

A. Ask for a longer answer with more detail B. Define the decision criteria (total cost of ownership, implementation risk, support terms), request a criteria-by-criteria comparison quoting each proposal, and require the reasoning before the recommendation C. Ask Claude to role-play as the CFO D. Ask for a one-word verdict to keep it crisp

Approach: Analysis tasks are graded on traceability. Named criteria turn "better" into something checkable; quoted evidence anchors every judgment to the source; reasoning-before-verdict prevents rationalizing a snap answer. This is the analysis column of the task-type table, applied. Why the others fail: A adds length, not basis — more unanchored words. C changes the persona, not the evidentiary structure; a role doesn't create criteria. D maximizes the exact weakness — a verdict with no visible support. Answer: B

Exam traps

• Model-switching as a prompt fix. When output misses because the brief lacks context, audience, or format, no tier change helps. Escalate models for reasoning depth, never for missing specification.
• Emphasis instead of specification. "Be careful," "be thorough," caps, exclamation marks — exhortation options are always wrong. The correct option adds a rule, an example, or a fact.
• Restarting when in-chat iteration would do — and its mirror image, grinding on in a chat polluted by contradictory corrections when the reset move is due. The tell for reset: iterations are making output worse.
• "Regenerate and pick the best" as a reliability strategy. Re-rolling doesn't fix a prompt that can't specify the target; consistency comes from examples and specifications, not sampling.
• Applying one task type's strategy to another: demanding citations in a brainstorm, letting a research brief run on training knowledge, asking for a verdict with no criteria, revealing your preferred answer and calling it brainstorming.
• Decomposition options with no review between steps. Splitting without checkpoints keeps the labor and discards the benefit — the seams are where errors get caught.
• Tuning on one sample and shipping the template. Reusable prompts get tested against varied representative inputs; one document is an anecdote, not a test set.
• Treating role prompting as a capability switch. It shapes perspective and vocabulary; distractors claim it unlocks features, accuracy, or speed.