Builder Field Guide

How to write an AI build prompt you can test

Tell the coding assistant what a person should see after one change. Give it a small example you can check yourself. A four-order packing list shows how a plausible first draft can still send the wrong order to the packing bench.

Download the editable build brief

What would make “improve the packing list” testable?

At this fictional market stall, a packer opens All orders and sees every order in intake order. “Make the packing list easier” could mean bigger labels, a new sort order or fewer rows. The packer wants to know which paid orders still need packing. Here is the build prompt that follows from that answer.

Build prompt. Add a To pack view to the order list. Include an order if it is paid and still unpacked. Keep intake order. When nothing qualifies, show “No paid orders need packing.” Selecting All orders again should restore the full list. Switching views should leave each order’s payment and packing statuses as they were. Run the three examples below and report the actual result of each.

Four invented orders, in intake order
OrderPaymentPacking
R4PaidUnpacked
R1UnpaidUnpacked
R3PaidPacked
R2PaidUnpacked

R1 is the case a quick demo might miss. It is still unpacked, but nobody has paid for it. R4 and R2 also make order matter. Sorting by ID would quietly put R2 first.

Do both drafts produce the requested screen?

The first draft checks whether each order is unpacked. The revised draft also checks payment. Both run here with the same sample data and expected results. Each check starts fresh.

Acceptance checks written before running either draft
Starting point and actionExpected screen
Start with all four orders and select To pack.R4, R2
Start with just R1 and R3 and select To pack.No paid orders need packing.
Start with all four orders, select To pack, then select All orders.R4, R1, R3, R2

The table above lists the three checks you can try.

What should you send back when a check fails?

Turn a voice note into a test

If you already use Wispr Flow on Mac or Windows, place the cursor in a blank note. Hold your configured push-to-talk shortcut while you speak. Release it and leave the cursor there until the text appears. You can also type the same example in any text editor, without an account or microphone. Both routes give you a draft to review before you send it to a coding assistant. Wispr Flow’s desktop dictation instructions.

Fictional spoken note. “I picked To pack. I see R four, R one, R two. R one hasn’t been paid. Don’t mark it packed. It should disappear from this view. Keep R four first, then R two.”

Compare the draft with the sample table. Replace spoken numbers with the exact order IDs. Check the words unpaid and not packed against separate columns. A missing “not” would select the wrong orders. Write down both lists in screen order so someone else can spot an extra row or an accidental sort.

Reviewed prompt. Start with R4, R1, R3, R2 and select To pack. I saw R4, R1, R2. Show only paid orders that are not packed. Keep intake order and leave all statuses unchanged. I expect R4, R2. With R1 and R3 alone, I expect “No paid orders need packing.” Rerun the three checks and report what appeared.

For your own change, write down one ordinary example, one input that could fool a plausible implementation and one existing action people rely on. Run them through the changed product. If a check needs an environment you cannot access, name that check when you hand off the result. The fourth Builder Field Guide lesson helps you choose the first task to test.

GitHub’s task guidance recommends defining acceptance criteria when assigning a coding change. Cucumber’s reference describes checking what the user or another system can actually observe. GitHub’s guide to scoping coding tasks. Cucumber’s guidance on expected outcomes.

Share what you’re building

Send a working title and a short description for private review. Add a deck if you have one.

Open the short application

Applying creates no commitment.