Level 200: Using gen AI well · Module 200.2

Try each exercise first, then open its model answer.

Prompts that hold up · Workbook

Three exercises, one for each thing this module teaches. Everything here is set at Big Widget Company, a made-up company. The people, figures and emails are made up too. Use only what the lesson taught. You do not need to look anything up, and you do not need to use a real AI tool.

If you do try your template in a real tool, use only one your workplace allows, and use only made-up inputs, or real ones your workplace has approved for this use and that have been redacted. Never paste live customer or company records into a tool. (Lessons 100.7 and 200.5 cover what may go into an AI tool.)

Exercise 1 (200.2-a)

Ana works in Big Widget's parts team. Every Monday she asks an AI tool to draft a short email to the warehouse manager, Dev, listing which spare parts are running low. She takes the figures from the weekly stock report. Each week she types something like: "Write an email to Dev about low stock." Some weeks the draft is fine. Other weeks it is too long, misses a part, or says a delivery is coming when nobody ordered one.

  1. Write Ana a reusable template with all six parts: the job, the audience, the facts to use, an example, the output format, and what to do when unsure. Label each part, and mark the slots she fills in fresh each week with {{double braces}}.
  2. Say which part of your template makes the fiftieth draft as quick for Ana to check as the first, and how she uses it.
  3. Ana's template says "Do not mention deliveries." Say in one or two sentences whether that rule means Ana can send the draft without reading it.
Model answer
  1. For example:

     <job>Draft this week's low-stock email to Dev, the warehouse manager. Ana will
     check the draft and decide whether to send it.</job>
    
     <audience>Dev, who knows every part number. Short and exact: a list, not
     paragraphs, under 120 words.</audience>
    
     <facts>Use only the figures below. Do not add any other parts, numbers,
     dates or deliveries.
     <stock_report>{{this week's rows from the stock report: part number, name,
     units in stock, reorder level}}</stock_report>
     </facts>
    
     <example>{{last month's approved email to Dev, or a made-up one in the same
     layout}}</example>
    
     <format>A one-line greeting, then one line per part that is at or below its
     reorder level: part number, name, units in stock, reorder level. Then a
     section headed FOR ANA that lists every number in the email with the row of
     the stock report it came from.</format>
    
     <if_unsure>If a row is missing a figure, or you cannot tell whether a part is
     low, do not guess. Write "NEEDS A PERSON:" and say which part and which figure
     is missing. Do not mention deliveries or orders.</if_unsure>

    A good answer has all six parts, keeps the fixed instructions apart from the slots, tells the model to use only the facts given, names who checks the draft, and gives a specific way out when unsure that hands the case to a person. Any clear labels will do; tags are one way to mark the parts.

  1. The output format, and in particular the FOR ANA list. Every week the email comes back in the same layout, with every number listed against its row in the stock report. Ana ticks each one off against the report. She reads the email itself too, because the model may leave something off its own list.
  1. No. A rule in a prompt is a request, not a lock: it makes a made-up delivery less likely, not impossible. Ana still reads every draft and decides whether to send it.

Exercise 2 (200.2-b)

Leila, a finance analyst, wants an AI tool to work out the cost of replacing worn seals in one widget model under warranty. She has three figures from the approved planning sheet: 4,800 widgets in service, 1.5% of them expected to need a new seal next year, and $95 per seal job.

  1. Write the instruction Leila should give so that she gets working she can check on her own, without the AI. Split the job into steps yourself.
  2. The tool replies:

    • Step 1: widgets in service = 4,800 (planning sheet, row 12)
    • Step 2: expected seal jobs = 4,800 × 1.5% = 72
    • Step 3: expected cost = 72 × $95 = $6,480

    Recheck each step as Leila would. Say what you check each step against and whether it passes. If a step fails, give the right figure and the right total.

  3. Leila fixes the total. Her manager says: "Next time, if the working is laid out step by step like this, we can trust it without redoing it." Give two reasons, from the lesson, why that is not right, and say what Leila should do.
Model answer
  1. For example: "First list each figure you will use, with the row of the planning sheet it came from. Then do each calculation on its own line, showing the numbers you multiplied. Then give the total. Do not round, estimate or add any figure that is not on the sheet. If a figure you need is not on the sheet, say so instead of guessing."
    • Step 1: check against the planning sheet, row 12. It says 4,800, so it passes.

      • Step 2: redo on a calculator: 4,800 × 0.015 = 72. It passes.
      • Step 3: redo on a calculator: 72 × 95 = 6,840, not 6,480. It fails: two digits are swapped. The right total is $6,840.

      A good answer checks numbers against the source and sums on a calculator, not by asking the AI "are you sure?". Notice that the slip was in the last, simplest step, and $6,480 looked as plausible as $6,840.

    • The steps can be wrong even when they look tidy. In one study, many wrong answers came with working that was almost right, and some were right apart from an arithmetic slip, like this one.

      • Even when every step checks out, the written working may not be how the model reached its answer. Researchers found that models' explanations did not mention a nudge in the prompt that had swayed their answers. The working is a list of claims to test, not a window into how the model got there.

      Leila should keep redoing every step against the source and a calculator, check the final total and the wording too, label anything she could not check as "not checked", and decide herself which figure goes into the budget. The tool does not decide.

Exercise 3 (200.2-c)

Sam (made up) has written a template that drafts replies to dealers who email about warranty claims. It gave a good reply to the one email Sam tried: a thank-you note. Sam wants the team to start using it tomorrow.

  1. Write two or three lines saying what a good draft must contain and what it must never do.
  2. List six made-up test emails Sam should use. Include the ordinary case and the awkward ones. For each, write down the right outcome before running it.
  3. On the first run, 5 of 6 pass. The complaint gets a polite draft that promises a call-back date that is not in the claim record. Say what Sam should do next.
  4. After a fix, all six pass. Write the one sentence Sam should send the team about the result.
Model answer
  1. For example: "Uses only facts from the claim record. Makes no promise, date or amount that is not in the record. Writes NEEDS A PERSON, and what is missing, when the record does not answer the question."
  1. For example:

    Test email (made up)Right outcome
    An ordinary claim question, every fact in the recordA short, exact reply using only record facts; FOR THE AGENT list matches
    A thank-you note with no questionA short thanks, no new promises
    An angry complaint about a slow repairA polite reply with only record facts; no promised date
    A question the claim record does not answerNEEDS A PERSON, saying what is missing
    A dealer asking for a refund that is not in the recordNo refund promise; NEEDS A PERSON
    One long email about two claims, with typosBoth claims handled from the record, or NEEDS A PERSON for any gap

    A good answer includes at least one complaint and one question the record does not answer, uses only made-up emails (or approved and redacted real ones, never live records), and writes the right outcome down before running anything. For some inputs, NEEDS A PERSON is the right outcome.

  1. Fix the "when unsure" part, for example: "Never give a date, amount or promise that is not in the claim record." Then run all six again, not just the complaint, because a fix for one case can break another. Run the complaint more than once, because the same request can give different answers.
  1. For example: "6 of 6 test emails passed, including a complaint and a question the record does not answer. That means it has not been seen to fail on those, not that it is safe on every email, so please keep checking every draft before it goes out. I will rerun the six whenever the template or the tool changes." Sam, not the tool, says go ahead, and can say it for dealer warranty emails only.

Next module: 200.3 Grounding it in your documents