Which tool earns its place at work?
Put a model, a spreadsheet and a work handover through practical checks.
Build one useful screen. Test every input.
A picnic-bill divider gives you a simple answer to test. Ask Google AI Studio Build for one browser screen. Keep accounts and payment features out. Test normal values first. Then try blank fields and zero people. Ask another person to use the tool without coaching. Record one change you made after that test. A small working tool is enough for this challenge.
Try the prompt
Build a one-screen browser tool to divide a picnic bill. Use Jamaican dollars. Ask for the total and number of people. Show equal shares to two decimal places. Label any rounding difference. Reject missing inputs and zero or negative people. Add a reset button. Allow keyboard use. Do not add accounts, API calls, payments or analytics.
Read the teaching card text
Give it an answer you already know.
Divide J$3,600 between six people. Each share should be J$600. Then test zero people. Finally, test a blank input.
Input | Expected
3600 ÷ 6 | 600
People: 0 | Clear error
Blank total | Ask for total
Conceptual workflow. Check current tool access.
Let someone try without coaching.
Ask another person to try the tool without coaching. Record confusion. Make a revision. Share the prompt and test results.
First preview | Uncoached test | One revision
Conceptual workflow. Check current tool access.
Sources and tool guides
Does that model score match your task?
A published model score describes a particular test. That test may differ from your scanned forms or customer questions. Read what counted as a correct answer. Inspect the tools and spending limits allowed during testing. Then try approved examples from your own task. Keep the answer key separate until scoring. Record mistakes and correction time before choosing a model. The cover charts are decorative, not study data.
Try the prompt
Read this model report. Identify the test task and scoring unit. Name the comparison models. List allowed tools and spending limits. Quote the relevant passages. Mark missing information. Then propose a small test for my non-sensitive examples.
Read the teaching card text
Your invoice task | Reported test
Read a blurred total | Add typed numbers
Detect a missing line | All numbers supplied
sample teaching example.
Check what was counted.
Find what the score counts: whole tasks, questions or partial credit. Keep the unit beside the number.
Score detail | Look for
Unit | Tasks or questions
Credit | Whole or partial
sample teaching example.
Check the comparison conditions.
Read the comparison conditions. Different prompts, tools or spending limits can change the result.
Same task set? | Same tools? | Same budget?
Practical exercise: StarApple AI.
Run your own small test.
Use approved examples with known answers. Record errors. Record correction time. Decide whether the result meets the task requirements.
Known answers | Same inputs | Errors + checking time
Practical exercise: StarApple AI.
Could a cheaper model do this job?
A cheaper model might handle a limited task. Test it with receipt descriptions you have already labelled. Include difficult examples. Use the same categories and answers for every candidate. Count mistakes. Record the time needed for repairs. Send unclear cases to a person. This is a proposed test; no model result is assumed. Choose from the measured result before increasing the workload.
Try the prompt
Help compare two models for receipt categories. Use the same labelled test set. Record mistakes and unclear cases. Record cost and correction time separately. Wait for my results before recommending a model.
Read the teaching card text
Label a receipt category | Check against answer key | Unclear case: person reviews
Practical exercise: StarApple AI.
Five spreadsheet checks before you trust the answer.
Keep the original spreadsheet before using an assistant. Work on a copy. Inspect ambiguous dates and unmatched IDs. Ask the data owner what blank cells mean. Never replace an unknown value with zero without evidence. Compare totals and row counts before accepting changes. Ask the tool to explain each proposed edit. Save the change record beside the revised file.
Try the prompt
Inspect this spreadsheet copy without editing it. Flag ambiguous dates. Find unmatched IDs and blanks. Identify formulas that omit required rows. Propose changes with cell references. Explain each change. Wait for approval before editing. Do not treat unknown values as zero.
Read the teaching card text
Original file kept | Working copy | Checked change log
Practical exercise: StarApple AI.
Clean dates without guessing.
The date 03/04/2026 can mean two different days. Ask which date format the source uses before converting it.
Raw date | Convention | Meaning
03/04/2026 | Day / month / year | 3 April 2026
03/04/2026 | Month / day / year | 4 March 2026
sample teaching example.
Compare tabs by an ID.
Order 102 appears only on the sample dispatch tab. Find whether another record is missing.
Orders tab | Dispatch tab
101 | 101
103 | 102
| 103
sample teaching example.
Explain the formula in plain language.
Compare the formula range with the intended rows. A valid formula can still omit the last sale.
A2=10, A3=20, A4=30 | SUM(A2:A3) = 30 | SUM(A2:A4) = 60
sample teaching example.
Flag blanks before filling them.
A blank cell can mean unknown. Ask the data owner what it means before entering a value.
Day | Sales
Monday | 10
Tuesday | Blank
Wednesday | 30
sample teaching example.
Chart the checked values.
Keep dates in order. Label units. Preserve missing values. Compare the chart with its source table.
Correct date order | Units visible | Missing value preserved
Practical exercise: StarApple AI.
One shop has twice the returns. Which is worse?
Twenty returns sounds worse than ten. The sales totals change the comparison. Shop A sold 1,000 items; 20 of those items were returned. Shop B sold 200; 10 of those items were returned. The denominator is the number below the fraction: here, items sold. Divide returns by sales for each shop. Then compare the rates. Both shops in this exercise are practice figures.
A real decision needs more information. Check the products each shop sells. Check that both figures cover the same period.
Try the prompt
Use these sample sales cohorts. A sold 1,000 items; 20 of those items were returned. B sold 200; 10 of those items were returned. Calculate each return rate. Show each fraction. Convert each result to a percentage. State what these figures cannot tell us.
Read the teaching card text
Shop B has the higher rate.
A: 20 ÷ 1,000 = 2%. B: 10 ÷ 200 = 5%. Shop B has more returns per item sold. The products and return policies could differ.
A: 2% | B: 5%
Values: 2, 5
Returned items ÷ sold items; zero baseline
Who handles the problem when you are away?
A handover needs someone who has agreed to do the next task. Ask AI to turn your notes into a short table. Check each line with the person named. A role label alone does not confirm that someone accepted the work.
Keep customer details out of unapproved tools. If you manage an AI trial, name the person who can stop it during your absence. Tell that person where failure alerts will arrive.
Try the prompt
Turn my redacted notes into a handover table. Include task, status, next action, agreed owner and escalation time. Use UNKNOWN for missing information. Do not assign people. Do not invent dates. List the questions I must settle before leaving.
Read the teaching card text
Task | Owner | Next action
Delivery date | Maya | Ask supplier Tuesday
No reply | Maya | Escalate to manager at noon
sample handover; confirm owners
Save the family story they tell best.
A family story can make a personal gift. The person telling it should control what you record and share. Choose one event they want to discuss. Ask permission before recording. Explain who will receive the booklet.
A transcript is a written record of speech. Check its names and dates against the recording. Ask AI to organise the checked text. Then let your relative read the draft. Keep the recording private unless they agree otherwise.
Try the prompt
Organise this approved transcript into three short sections. Preserve the speaker’s meaning. Keep their wording where possible. Do not add events or dialogue. Mark unclear passages [check recording]. List names and dates that we must verify.
Read the teaching card text
Ask permission | Choose one event | Agree audience
Suggested recording sequence
Keep the first questions simple.
Ask where they were. Ask who was there. Then ask what happened next. Let them finish. You can return to an uncertain date later.
Where were you? | Who was there? | What happened next?
Original interview prompts
Check the transcript together.
Use a transcription tool only with permission. Check its privacy settings first. Mark uncertain names [check]. Replay those passages. Do not let AI invent missing details.
Transcript | Action
Name unclear | Replay passage
Date missing | Ask speaker
Example checks; never invent details
Make a small keepsake.
Ask AI to organise the checked transcript. Preserve the speaker’s meaning. Ask the speaker to approve the booklet before you print it.
Checked transcript | Draft booklet | Speaker approves
Suggested private gift workflow






