Level 200: Using gen AI well · Module 200.R

Try each exercise first, then open its model answer.

Gen AI in your role (role tracks) · Workbook

Two exercises, one per objective. Use made-up or approved and redacted examples only: never live customer or company records, real prompts or real outputs. Big Widget Company and its people are made up. The task card, the four kinds of check, the "drafts or acts" sort and the gate answers are this course's own practice, not a law or a standard.

Exercise 1 (200.R-a): write a task card for one recurring task

Pick one task you do every week or every month in your role. Use your track's row in the lesson, or write your own. If you use AI only at home, use the home track.

  1. Write the four parts of the task card:
    • Prompt template: the job, the audience, the facts to use, an example, the output format, and what to do when unsure. Mark the blanks that change each time, like [CLAIM].
    • Grounding source: the one document or system the answer must come from, and its owner (a person or a named role, not "the team").
    • Main check: citation check, numbers check, tone check or legal review. Say why it is the main one: what is the costliest mistake it would catch? Name any second check.
    • Data rule: what may go in, what must never go in, and what you do instead.
  2. Test the template on three made-up inputs, including one awkward one. Write one line on what each test showed.
  3. Read one other track's row. Name its main check and say in one or two sentences why it differs from yours.
Model answer

A good answer has all four parts, each specific enough that a colleague could run the task from the card alone. For example (made up), Lee, a finance analyst at Big Widget Company:

  • Prompt template: "You are drafting the month-end variance note for the [MONTH] management pack. Use only the ledger export pasted below. For each line in [LINES] whose variance is over [THRESHOLD], give the figure, the budget and one sentence on the likely cause, as a bullet list for department heads. Do not estimate any figure the export does not contain. If you cannot explain a variance from the export, write 'Needs a person: cause not in the data'. Example bullet: [EXAMPLE]."
  • Grounding source: the month-end ledger export from the finance system. Owner: the financial controller (a named role).
  • Main check: numbers check. Lee recomputes every figure in the draft from the ledger export, with the finance system or a calculator, before the note goes out. It is the main check because a wrong figure in a management pack can drive a wrong decision, and the wording cannot hide that. A tone check comes second, because department heads read the note.
  • Data rule: ledger figures for the month may go into the approved company assistant. Unpublished results never go into any other tool, and staff names never go in. Instead, Lee uses the cost-centre code.
  • Tests: a normal month; a month with one large variance; and an awkward one, a line with no budget, where the template should say "Needs a person" rather than make up a cause.
  • Other track: brand and communications. Its main check is tone, because the costliest mistake in a public statement is how it reads to the public, and a public-facing bot's words are published in the company's name. Finance's costliest mistake is a wrong number, so its main check is different.

Weak answers to look for:

  • The grounding source is "the internet", "what the AI knows" or "the shared drive", or the owner is "the team".
  • The check is "ask the AI to double-check" or "make the prompt better". Asking the same model is not a check, and a better prompt is not a substitute for one.
  • The data rule says only "be careful" or "no personal data", with nothing on what is allowed or what to do instead.
  • The card would fit any job unchanged. The four parts are the same in every job; what goes in them is not.

Exercise 2 (200.R-b): sort your team's tools and name who approves

List every AI tool your team uses, or use the made-up list below if you cannot share yours.

Tool (made up)What the team uses it for
Company chat assistantDraft supplier emails from pasted text; a person sends them
Meeting-notes tool, linked to the calendar and emailSummarize meetings; settings allow it to email the summary to attendees by itself
Purchasing assistantSuggest reorders; it can also place orders under a set amount
Report summarizerTurn the weekly safety log into a summary a person reads

For each tool:

  1. Write what it can do, not only what you use it for. Does it send, change, buy or publish?
  2. Sort it: drafts or acts.
  3. For each action, give a gate answer (allow, constrain, hold or deny) and name who approves it: a person or a named role. If you do not know, write "not known yet: ask [name]".
  4. Say where the gate is enforced: in the tool's settings or permissions, not in the prompt.
Model answer
ToolCan doDrafts or actsGate answer and approver
Company chat assistantRead pasted text; no sendDraftsNo approver needed: a person sends
Meeting-notes toolRead the calendar and email; send email by itselfActs (sends)Hold: auto-send off in the settings; the meeting owner approves each send. If the meeting owner is not settled: "not known yet: ask the IT change owner", and held until then
Purchasing assistantSuggest orders and place ordersActs (buys)Hold: ordering off in its permissions; a named buyer approves each order. If the team wants small orders to run by themselves, that is a constrain answer, with the limit written down and a named approver for the rule
Report summarizerRead the log; write a summary for a personDraftsNo approver needed for the summary; a person reads every serious report

A good answer:

  • sorts by what the tool can do. The meeting-notes tool is used to summarize, but because it can send by itself, it acts. OWASP's example is the same: a tool that reads a mailbox to summarize email may only need to read, so it should not also be able to send or delete messages (the lesson's source s4).
  • names a person or a named role for every action, never "a human in the loop" or "the AI asks first". A "not known yet" with the person to ask is a good answer; until someone is named, the action is held.
  • puts the gate in settings or permissions. "Always ask before sending" in a prompt is not a gate.
  • treats anything not listed as denied, and any tool whose abilities are unclear as one that acts, with every action held until its owner says otherwise.
  • notes that changing a tool's set-up (switching on auto-send, adding a connector, widening access) is an action too: it needs a named approver, and the old yes does not cover the changed tool.

Anything you bring to 200.P from this workbook holds decisions and made-up examples only: no real prompts, outputs, customer records or personal data.

Next module: 200.P Project: retrofit a rule on a tool you use