Free CCAO-F flashcards

When there is not enough time to check every sentence of a Claude draft, which kinds of claim should you verify first?

Claude can produce fluent text that still contains a wrong date, figure or quotation. Those specific claims, such as names, numbers, dates and words attributed to a person, are what readers act on and what is hardest to retract once published. Checking them against the approved primary document gives the most protection for limited time, while general background and stylistic choices carry far less risk.

Does a confidence percentage that Claude gives for its own work show that its figures are accurate?

No. When Claude states a confidence percentage it is generating a plausible-sounding answer, not measuring how often its output matches the source. The only way to know the figures are right is to compare them with the original documents, and passing the rating on to decision makers replaces that evidence with a number that means little and may give false assurance.

Common misconceptionAsking Claude to list the figures it feels least sure about does not target your checking well, because its own sense of uncertainty does not reliably identify where its errors are.

Which kinds of HR task belong on a shortlist of Claude use cases, and which must stay with a person?

Value comes from matching tasks to what Claude does well, such as drafting and summarising, while keeping accountability with people where it must sit. Hiring, disciplinary and probation outcomes affect individuals and can carry bias, so a person decides them. Drafting adverts for a recruiter to review, or theming anonymised feedback, leaves every decision with staff.

When using Claude to analyse conflicting stakeholder notes for requirements, what should Claude do and what must stay with the stakeholders?

Claude compares many inputs quickly and exposes contradictions, gaps and assumptions, but it cannot know which competing requirement the business wants, so the stakeholders resolve the conflicts. Asking Claude to link each restated requirement to its source note makes the output checkable, which is a stronger safeguard than any rating Claude gives itself.

Common misconceptionA completeness score Claude gives its own requirements is not evidence that they are complete. Only checking against the source notes and the stakeholders' answers shows whether the requirements are right.

How do you decide whether a task involving staff is an appropriate use of Claude, rather than following a blanket rule?

Appropriateness depends on the task, the data and the policy, not on whether the work touches staff. Drafting internal material that holds no personal data, with human review, is the kind of work Claude suits. The review step is what makes the draft safe to release, because a subject-matter expert catches any step Claude has described wrongly before people rely on it, so it is kept rather than skipped.

When Claude helps with a hiring shortlist, which part of the work can it take on, and which must stay with a person?

Shortlisting is a decision about a person, so it needs a human who can be held accountable and who can notice when the output misreads someone. Summarising applications against stated criteria is drafting and analysis work that Claude does well, and it speeds the recruiter up without handing over the decision. Where a policy names who decides, that settles the line between a supporting use and an automated decision.

Common misconceptionApplying objective criteria more consistently does not make an automated rejection acceptable. A rejection is still a decision about a person, and Claude can misread or unevenly weigh applications in ways nobody catches if no human looks.

Your organisation has known rules for a piece of writing, such as a word limit or banned language. How do you get Claude to follow them?

State the rules in the prompt as explicit constraints. Claude cannot follow a style guide it has not been shown, so it falls back on general conventions for the genre. Writing the limits and prohibitions into the prompt turns unstated expectations into instructions, which is far more reliable than asking for a vaguer quality like persuasiveness or fixing every draft by hand afterwards.

Common misconceptionAsking Claude to rate its draft against best practice is not a compliance check. It has not been told your organisation's rules, so it measures generic practice, and a self-rating is not evidence the draft meets the guide.

Claude's response is accurate but pitched at the wrong level for its readers. What should you add to the prompt to fix the register?

Name the intended audience and the reading level they need. Claude infers tone and complexity from whatever the prompt gives it, and with no audience named it tends to mirror the source material, so a legal policy produces legal phrasing. Telling Claude who will read the output and how plainly it must be written changes the register directly, which rerunning the same prompt cannot do.

Common misconceptionA more capable model will not write more clearly for your readers on its own. No model can guess an audience the prompt leaves out, so switching costs more without supplying the missing information.