StarApple AI Fieldnotes

Would you remember how to do it?

Try a recall test, catch a mistake and set a limit before buying.

A marked-up proof sheet has three separate errors circled in red.

Can you catch three AI mistakes?

This sample brief contains three errors. Make your own corrections before asking AI. Calculate the percentage from the source counts. Calculate energy from power and time. Power measures the rate of energy use. Energy measures the total used over time. Identify the claim with no supplied source. Explain every correction with arithmetic or evidence.

Try the prompt

Inspect this sample brief against the supplied facts. Source: 40 reports, 10 with errors; power 2 kW for 3 hours. Brief: 10% errors; energy 6 kW; a national study proves universal applicability. Show the arithmetic. Identify unsupported claims. Do not invent a study.

Read the teaching card text

Audit this made-up AI brief.
“Only 10% of reports had errors. The pilot used 6 kW of energy. A national study proves universal applicability.” Find three errors.
Percentage | Energy unit | Evidence for wider claim
Practical exercise: StarApple AI.

Show your corrected version.
Ten out of 40 is 25%. Energy is 6 kWh. No national study was supplied. The wider claim is unverified.
10 ÷ 40 × 100 = 25% | 2 kW × 3 h = 6 kWh | National claim: no supplied source
sample teaching example.

Individual wooden letterpress sorts break away from a line of metal type in a shaft of warm light.

Why does AI count text differently?

Token counts can differ from word counts. A token can be a word, part of a word or punctuation. The model and its encoding determine the split. An encoding is the set of rules that maps text to token IDs. Try a short sentence in the official tokenizer. Change a capital letter. Compare the split. A token ID is a number that identifies a text token. For software-based requests, inspect actual usage records before estimating costs.

Try the prompt

Explain words and tokens in plain language. Define encoding and token ID. Do not invent exact token boundaries or counts. Name the tokenizer for my chosen model. I will use it to test a sentence.

Read the teaching card text

Why does AI count text differently?
A token can be a word, part of a word or punctuation. The model determines the split.
Text you enter | Model-specific token split | Model processes token IDs
Conceptual workflow. Check current tool access.

Sources and tool guides
Two people exchange a stock sheet over a market counter filled with spices, beans and packaged goods.

Five shop tasks you can try today.

Start with one written shop task. A shift handover or product description can be a useful test. Use sample notes for the first attempt. Give the tool only approved facts. Inspect its draft before anyone uses it. The shop owner must approve prices and customer promises. Save the checked version where the next shift can find it.

Try the prompt

Use only these approved shop notes. Draft a short handover. Separate completed tasks from open questions. Mark unknown stock quantities. Do not invent prices or customer promises. Remove customer details that identify a person.

Read the teaching card text

Approved shop facts | AI draft | Owner checks
Practical exercise: StarApple AI.

Turn stock notes into questions.
sample note: red mugs low; blue mugs counted yesterday. Ask which facts need inspection before a reorder. Do not invent quantities.
Item | Next check
Red mugs | Count stock
Blue mugs | Check sales since count
sample teaching example.

Draft an answer from your policy.
Use the written returns policy to draft answers. Request the policy sentence behind each answer.
Approved returns policy | Draft FAQ | Check each promise
Practical exercise: StarApple AI.

Write product copy from facts.
Supply material, size and care instructions. Remove claims that the facts do not justify.
Source fact | Allowed copy | Unsupported claim
Cotton shirt, size M | Cotton shirt, size M | Stain-proof
sample teaching example.

Prepare the next shift.
Turn approved notes into a handover. Give unresolved problems a separate line.
Task | Status
Collect order | Ready
Confirm repair | Waiting for owner
sample teaching example.

Read review themes carefully.
Group public reviews by issue. Compare the summary with the original wording. Remove names before sharing it.
Review theme | Check
Late delivery | Dates and context
Friendly staff | Actual review wording
sample teaching example.

A student sets a tablet face down beside an open notebook of geometry problems.

Will you remember it without AI?

A Turkish secondary-school trial compared two AI maths tutors with a control group. The control group practised without an AI tutor. Both AI designs improved scores during assisted practice. A percentage point measures the difference between two percentage scores. Results differed when students worked without AI. The study could not establish an exam-score difference between the teacher-designed tutor and the control group. This does not prove identical results. A short prompt does not reproduce that system. Try a problem first. Ask for one hint. Then solve a fresh problem without AI.

Try the prompt

First ask me to attempt this maths problem. Give one small hint about the next step. Wait for my reply. Do not give the final answer. When I finish, give me a different problem for unaided practice.

Read the teaching card text

Will you remember it without AI?
Maths trial effects versus control, in percentage points. Base: standard chat. Tutor: teacher-designed hints. Tutor exam effect: not statistically distinguishable from zero.
Practice: Base | Practice: Tutor | Exam: Base | Exam: Tutor
Values: 13.7, 36.1, -5.4, -0.4
Bastani et al., Table 1. Percentage-point effects versus control. Tutor exam effect: not statistically distinguishable from zero.

Sources and tool guides
A red pencil connects three dots along the diagonal of a hand-drawn nine-square puzzle.

Can you spot three useful checks?

Find a straight line of three useful checks. Explain each choice before reading the answer card. This static puzzle has one intended winning line. The other squares offer shortcuts that do not establish whether a claim is correct. Try the puzzle with a colleague. For real work, use all the checks the task needs. One completed line does not finish a factual review.

Try the prompt

I chose A, E and I in this puzzle. Ask why opening a source helps. Then ask why dates matter. Finally, ask why I should recalculate a total. Challenge unclear explanations. Do not add new rules.

Read the teaching card text

The winning line is A, E, I.
Open the source. Inspect its date. Recalculate the total. These actions examine evidence.
A: Source | E: Date | I: Calculation
Practical exercise: StarApple AI.

A hand colours a tomato-vine drawing in a garden beside the real plant.

Make a colouring book they will recognise.

A familiar hobby gives a colouring book a personal theme. Choose six original scenes. Ask an image tool for one page first. Inspect the outlines. Print a test copy before creating the remaining pages. Small spaces can be hard to colour even when the screen preview looks good. Get permission before including a recognisable person. Keep private photographs out of public prompts.

Try the prompt

Create one original colouring page about watering a tomato plant. Use black outlines on white. Make the spaces large. Show one main subject. Give the watering can a closed handle. Leave print margins. Do not add text, shading or copyrighted characters.

Read the teaching card text

Gardening: sow seeds | Water the seedlings | Pick ripe tomatoes
Conceptual workflow. Check current tool access.

Make one test page first.
Ask ChatGPT Images for black outlines on white. Request large spaces and one main subject. Exclude copyrighted characters.
Black outlines | Large open spaces | One main subject
Conceptual workflow. Check current tool access.

Inspect the awkward details.
Inspect fingers and handles. Find lines that should join. Request a revision before creating the other pages.
Image problem | Requested repair
Handle has a gap | Join the handle outline
Petal lines overlap | Separate the petal outlines
sample teaching example.

Print before you give it.
Inspect the paper size and margins. Print one page. Use the successful test as a reference for the remaining five.
One image | One print test | Six checked pages
Conceptual workflow. Check current tool access.

Sources and tool guides
An open fountain pen lies crosswise over the blank signature area on a purchase order.

What would make you cancel the purchase?

Write your stopping condition before asking AI to defend a purchase. A failed trial task could be one condition. An unmet privacy rule could be another. Ask the tool to find evidence against the purchase. Open the sources yourself. Keep the original condition visible while comparing options. A persuasive explanation should not change your acceptance test without a reason.

Try the prompt

I am considering this decision: [decision]. Ask which facts would make it unsuitable. Ask what evidence I have. Separate verified facts from assumptions. Suggest one small test before I commit.

Read the teaching card text

Trial task | Pause if | Check
Summarise five orders | Any order is omitted | Count all five IDs
sample teaching example.