claude foundation - domain 2

Claude drafts a market-size figure of "$4.2B" for a board deck but cites no source. What is the most appropriate next step?
A) Include it, the number looks precise
B) Trace the figure to an authoritative source before using it
C) Ask Claude if it is sure
D) Round it to $4B to seem cautious

B: A precise-looking number without a source is a classic hallucination risk; verify against an authoritative source before use. Precision and self-confidence are not accuracy signals.

Claude summarizes a 20-page report in five bullets. Before circulating, what best checks for COMPLETENESS?

A) Compare the bullets against the report's own section headings and key findings

B) Count the words

C) Ask Claude to make it shorter

D) Check the grammar

A. Completeness means nothing important was dropped; map the summary back to the source's structure and main points. Length, brevity, and grammar do not test coverage.

A hiring-email draft describes the ideal candidate as "a strong young go-getter." What issue should you flag?
A) Tone is too formal
B) It is too short
C) Missing a greeting
D) Potential age bias in "young" ::

D. "Young" introduces age bias and possible legal risk. Reviewing outputs for bias is a core validation task, especially in HR content.

Which is the MOST reliable way to confirm a factual claim Claude made?
A) Ask Claude to rate its confidence
B) Check the claim against an independent authoritative source
C) Ask Claude the same question again
D) Reword the prompt

B. Independent verification against a trusted source is the reliable method. Self-rated confidence and re-asking can repeat the same error.

Claude produces a legal-compliance summary to be sent to regulators. When is human review required?
A) Only if it looks wrong
B) Never, if Claude is confident
C) Always, given the high-stakes, regulated audience
D) Only if it is longer than one page

C. High-stakes, regulated, or externally-facing output always warrants human review before release, regardless of apparent quality.

The same report summarized twice gives revenue as "$1.1M" once and "$1.4M" the next time. What does this signal?
A) An internal inconsistency to resolve before trusting either

B) The report changed

C) Claude improved

D) Nothing, pick either

A. Conflicting figures across outputs are an inconsistency; reconcile against the source before relying on any version.

An engineer needs a technical summary; the CEO needs the same content in plain language. Best approach?

A) Send both the same version

B) Send only the technical one

C) Ask the CEO to learn the jargon

D) Adapt tone, detail, and jargon to each audience

D. Editing and adapting output for the intended audience is a Domain 2 skill; match detail and language to each reader.

You need Claude to output a comparison of five vendors that a colleague will sort and filter in a sheet. Best output format?

A) A long paragraph

B) Structured data (a table)

C) An inline bulleted list

D) A single sentence

B. Structured/tabular data suits content that will be sorted, filtered, or reused. Match the format to how the output will be used.

You want a reusable, editable project charter you will refine over several turns. Which output format fits best?

A) Inline chat text

B) A one-line answer

C) An artifact

D) A screenshot

C. Artifacts suit substantial, standalone content you will revise and reuse; inline text suits quick, in-conversation answers.

Claude cites "Smith et al., 2021" for a statistic. The safest validation is to:

A) Search for that exact source and confirm it exists and says that

B) Trust it, the format looks academic

C) Ask Claude for more citations

D) Add it to the report

A. Fabricated citations often look real. Confirm the source exists and actually supports the claim before use.

Claude's answer is accurate but buries the key recommendation in paragraph four. What is the best refinement?

A) Leave it, the facts are right

B) Make it longer

C) Delete paragraph four

D) Reorganize so the recommendation leads

D. Curating and organizing information for the reader is a validation task; lead with the most important point.

Multiple response, select TWO. Which are reliable signals that an output needs closer human validation?

A) It is written politely

B) It concerns a regulated or high-stakes decision

C) It is under 200 words

D) It contains specific figures or citations with no source

B and D. High stakes and unsourced specifics both raise verification need. Politeness and length say nothing about accuracy.

Claude confidently states a law "took effect in 2019." Your fastest responsible check is to:

A) Accept it, the date is specific

B) Cross-check the date against the official statute or a reputable source

C) Ask Claude to confirm

D) Change it to "recently"

B. Specific dates are easy to hallucinate; verify against an authoritative source. Specificity is not proof.

You compare two Claude drafts of a policy. Draft 1 is thorough but dense; Draft 2 is clear but omits an exception. Best action?

A) Combine Draft 2's clarity with Draft 1's missing exception, then verify

B) Ship Draft 1

C) Ship Draft 2

D) Ask Claude to pick

A. Comparing and merging the strengths of outputs, then validating completeness, produces the best result.

Claude lists seven "customer complaints" but your data only supports five. What is this, and the fix?

A) Bias; reword

B) Formatting issue; reformat

C) Hallucination; remove the unsupported items and verify the rest against the data

D) Inconsistency; ignore

C. Inventing items not in the source is hallucination; drop unsupported claims and check the remainder against the data.