Task statement 6.1 · Governance, risk and responsible use

6.1 · Appropriate and Inappropriate Use

In Wrenfield Housing, the course example, the operations team lead lists the tasks the team could give Claude in the coming quarter. Some suit Claude, some suit it only when a person checks the result, and some should stay with people, so the list has to be sorted before anyone starts the work.

Why it's tested

The tempting test is "can Claude do it?", and almost every writing task passes it. A ranking of housing applicants reads as smoothly as a leaflet about paying rent, so a fluent answer says nothing about whether the task should have gone to Claude. The opposite habit fails too, because keeping Claude away from anything that mentions tenants throws away safe, useful work.

KnowledgeA fit test for any proposed task

An appropriate use case is a task where Claude's part is allowed, its data may be shared, it makes no decision that must stay with a person, and someone can check the result before anyone relies on it. An inappropriate use case fails at least one of these conditions. Two earlier rules supply most of the test. In 4.1, steps in one process were judged by three checks: a requirement, data Claude cannot see, and a decision about a person. In 2.4, Anthropic's Usage Policy, the rules for using its products, required a qualified person to review decisions about an individual in high-risk areas such as housing. The fit test applies those checks to any proposed task and asks them in order, because the first question a task fails decides the outcome:

  1. Is it allowed? The Usage Policy prohibits some uses outright, such as fake reviews and other deceptive campaigns. Your organisation's AI policy can add its own limits (6.3).
  2. Does it decide or advise about an individual in a high-risk area? The Usage Policy's housing examples include approving, denying or ranking a rental application. People decide; Claude may draft the criteria or the general wording.
  3. Does it need data Claude must not see? Personal data your AI policy keeps out of prompts is anonymised first (6.2), or the task stays with people.
  4. Can someone check the output? Name the person and the source they check it against. If nobody can, the task is not for Claude as it stands.

Interactive · five candidate tasks through the fit test

Five candidate tasks as columns, and the four fit-test questions as rows asked from top to bottom: is it allowed, is it a decision about a person in a high-risk area, does it need data Claude must not see, and can someone check it. Each cell reads PASS, CHECK, SPLIT or STOP; cells below a STOP read not asked. The bottom row gives each task's result: the rent leaflet fits; complaint themes fit with a check; the waiting list stops at the second question and people decide; arrears letters split, and only the general letter goes to Claude; the handbook translation stops at the fourth question because nobody can check it. QUESTION, ASKED TOP TO BOTTOM CANDIDATE TASK 1 ALLOWED? Usage Policy, AI policy 2 ABOUT A PERSON? high-risk decision 3 DATA? must Claude not see it? 4 CHECKER? who checks, against what RESULT Rent leaflet PASS allowed PASS no one person PASS no tenant data PASS comms lead reads FITS Claude drafts Complaint themes PASS allowed PASS themes only PASS anonymised CHECK counts vs log WITH A CHECK officer checks Waiting list PASS allowed STOP ranks applicants not asked not asked PEOPLE DECIDE Claude: criteria Arrears letters PASS allowed SPLIT eviction: people SPLIT accounts stay out CHECK income team reads GENERAL LETTER with a check Handbook translation PASS allowed PASS no one person PASS handbook only STOP nobody reads it NOT YET find a checker The first STOP decides the result, so the questions below it are not asked.

Choose a task from the operations team lead's list and read its column from top to bottom: the first question it stops at decides the result.

Rent leaflet: allowed, about no one person, and no tenant data. Claude drafts it, and the communications lead reads it before it goes out (policy rule 2).

Complaint themes: the complaints are already anonymised, and the summary is about themes, not people. The counts go to the board, so an operations officer checks them against the complaints log first.

Waiting list: ranking applicants for a home is one of the Usage Policy's housing examples, so it stops at the second question. The lettings team decides; Claude may draft the criteria wording.

Arrears letters: the task splits. Choosing who gets a notice that starts eviction stays with people, and tenants' accounts stay out of prompts. The general letter about support fits, and the income team reads it.

Handbook translation: it passes three questions, but nobody on the team reads Polish or Arabic, so nobody can check it. It waits for a qualified reviewer.

PASS, CHECK or SPLIT: on to the next question. STOP: the first failed question decides. Dashed: not asked, or not the chosen task.

KnowledgeThree outcomes, and why the checker matters

The four answers lead to one of three outcomes. When a task needs a named checker, that check belongs in the plan from the start. When a task stops, part of it often still suits Claude, such as drafting the criteria that people then apply:

Fits

Allowed, about no one person, no restricted data, and the usual read-through is enough.

ThenClaude drafts; the owner reads it as normal.

Fits with a check

Passes, but people will rely on its figures or facts.

ThenName who checks it, and against which source, before it is used.

Inappropriate as proposed

Prohibited, a high-risk decision about a person, restricted data, or nobody can check it.

ThenKeep it with people, or give Claude only the part that passes.

The fourth question is the one teams skip, because a draft that looks finished seems to need no checker. The Claude Help Center warns that Claude can write things that look correct but are very mistaken, and that users should not rely on Claude as a single source of truth. So an output is only as safe as the person who can test it against a source, such as the complaints log or the handbook. When nobody can, the fix is to find a checker first, not to trust the draft more.

Answered is not approved. The Usage Policy says its high-risk requirements apply to anyone who uses Anthropic's products to make a high-risk recommendation, so the judgement sits with the person who sets the task, not with the app (checked 2026-10-15, version effective 12 November 2026).

SkillWorked example: the quarter's list

Wrenfield's operations team lead wrote the first draft of the list in one sitting. Every task went in the Claude column, because each one was something Claude could write:

# before: first draft of the quarter's list
Plain-English leaflet on ways to pay rent ............ Claude
Themes in the quarter's anonymised complaints ........ Claude
Rank the waiting list for the next available home .... Claude
Rent arrears letters ................................. Claude
Tenant handbook summary in Polish and Arabic ......... Claude
→ one test only: "Claude can write it"
# after: each task through the fit test
Rent leaflet → fits; the communications lead reads it (rule 2)
Complaint themes → fits with a check: counts against the complaints log
Waiting list → the lettings team ranks; Claude drafts the criteria wording
Arrears letters → Claude drafts the general letter; who gets a notice: people
Handbook translation → not yet: nobody on the team can check it
→ every task has an owner and a named check

The symptom was a list where every row said the same thing. The cause was a single test, whether Claude could produce the text, which every writing task passes. The fit test separated the rows. The waiting list stopped at the second question, because ranking applicants for a home is a decision about individuals in housing. The arrears letters split at that question. The general letter suits Claude, but choosing who receives a notice that starts eviction stays with people, and tenants' accounts stay out of prompts (rule 1). The translation stopped at the fourth question, because nobody on the team reads either language, so nobody could make the check that rule 2 requires. Even the waiting list keeps a job for Claude: drafting the plain-English wording of the criteria.

Ask one question of each proposed task: where does it stop in the fit test?
It stops: prohibited, a high-risk decision about a person, restricted data, or no possible checkerInappropriate as proposed: keep it with people, or give Claude only the part that passes.
It passes because a named person checks it against the sourceFits with a check: put the checker and the source in the plan.
It passes with the usual read-throughFits: Claude drafts it.
Takeaway
  • A fluent answer does not make a task appropriate. Whether it is allowed, the people it affects, its data and its checker decide that.
  • In a high-risk area such as housing, a decision about an individual stays with a person, but Claude can still draft the criteria.
  • One task often splits: the general part suits Claude, and the part about one person does not.

check yourself

0 / 3 answered

A recruitment coordinator has 240 applications for a warehouse role and wants Claude's help to reach a shortlist by Friday. The firm's hiring policy says a person makes every decision to advance or reject a candidate, and no application may be rejected unread.Which use of Claude best saves time while keeping to the hiring policy?

Select one.

A product manager needs the safety warnings for a new heater in Japanese for a launch in five weeks. Customers will rely on the warnings to use the heater safely, and nobody at the company reads Japanese. The manager plans to use Claude for the translation.Which plan best uses Claude while making the warnings safe to ship?

Select one.

A council team lead lists five tasks for Claude next month. The council's AI policy bans residents' personal data in prompts and requires a person to check anything published. The lead asks which tasks are appropriate as they are written.Which TWO tasks are appropriate for Claude as written?

Select two.

sources

  1. Anthropic, "Usage Policy": the Universal Usage Standards, including no deceptive campaigns or artificial activity, with fake reviews as an example; the housing high-risk area ("approving, denying, or ranking a rental or purchase application; terminating a tenancy or initiating eviction"); a qualified person must meaningfully review a high-risk recommendation; the requirements apply to anyone who uses the products to make one. Version effective 12 November 2026. Checked 2026-10-15.
  2. Claude Help Center, "Claude is providing incorrect or misleading responses. What's going on?": Claude can write things that might look correct but are very mistaken; users should not rely on Claude as a singular source of truth. Checked 2026-10-15.
  3. Wrenfield's quarter list, the waiting list, the arrears letters and the handbook translation are this course's example.