Fix it

1. Review your card 08 results. Start these maker checks during Sprint 1; do not leave the full checklist until this short repair block. Your tester ran your checks, not the full audit. Add failures or unrun items honestly to your run log (card 05).
2. Pick one failure from your run log. Most serious first: wrong content, then anything on your red lines or "must never" list, then the learning design, then access, then looks.
3. Ask for one change. Use a prompt below. One change at a time, so you know what fixed it and what broke.
4. Every fix prompt ends: "Keep everything else unchanged. Explain what you changed and how I'd check it."
5. Check it yourself. Re-run the check that failed, then the checks that passed before. A fix can break something that worked.
6. Save a version each time a fix passes: -v2, -v3. Keep the old versions: going back is not always as simple as pressing undo.
   • Level 3 ■ Black: commit, with the check that now passes in the message. If Git isn't installed, save the numbered copies (card 12).
   • Level 1 ● Blue in Canva: Your save point is the version number. Note the number each time a fix passes.

"A tidy-up of the code quietly removed the button for loading older transactions, leaving most of the history unreachable."

Alex Gray, I Gave ChatGPT a Week to Build My App

Why ask for the explanation? In How AI assistance impacts the formation of coding skills (Anthropic, 29 January 2026, anthropic.com/research/AI-assistance-coding-skills), 52 mostly junior software engineers learned a Python library they had not used before. On a quiz straight afterwards, those who worked with AI scored 50% on average, against 67% for those who coded by hand. Within the AI group, those who asked it for explanations averaged 65% or more; those who handed the work over averaged under 40%. The authors say this second finding is an association, not proof of cause. And they were engineers, not teachers.

First finish your level's evidence on track card 10, 11 or 12. Then choose a useful task from card 14. Write the new success check first, make one change at a time, and rerun the original checks after each one. A peer handoff is usually more useful than an unrelated second project.

Think (AI inside the tool) is beyond this session: see card 18.

Fix a failed check

Success check [number] failed. When I [what I did], I expected [what should happen], but [what happened]. Fix only this.
Keep everything else unchanged. Explain what you changed and how I'd check it.

Correct a fact

[Question 4 / the formula for X / the label on Y] is wrong. The correct version is [correct version], because [reason]. Change only that.
Keep everything else unchanged. Explain what you changed and how I'd check it.

Remove decoration

Remove everything decorative: confetti, emoji, sound effects, animation that does not show change, stock images, points, streaks, badges, leaderboards and timers. Keep the layout plain and calm.
Keep everything else unchanged. Explain what you changed and how I'd check it.

Add a prediction step

Before students see [the answer / the result / the simulation run], add a step where they predict what will happen and commit to it by choosing or typing. Then show the result next to their prediction and ask them to explain any difference.
Keep everything else unchanged. Explain what you changed and how I'd check it.

Make feedback explain why

Rewrite the feedback. For every answer, right or wrong, say in one or two sentences a [age]-year-old can read why it is right or wrong, and what to do next. For [common wrong answer], name the misconception. No praise-only feedback.
Keep everything else unchanged. Explain what you changed and how I'd check it.

Simplify the controls

Reduce the controls to only these: [control 1], [control 2]. Remove the rest. Each change should show its effect straight away.
Keep everything else unchanged. Explain what you changed and how I'd check it.

Fix contrast and colour

Make all text contrast at least 4.5:1 against its background. Wherever colour signals something, such as right or wrong, add words too. List each text and background colour pair with its contrast ratio.
Keep everything else unchanged. Explain what you changed and how I'd check it.

Fix keyboard use

Make every control usable with the keyboard alone (Tab, Shift+Tab, Enter, Space, arrow keys), in a sensible order, with a clearly visible focus outline. Give every button and input a label a screen reader can read.
Keep everything else unchanged. Explain what you changed and how I'd check it.

Make it work at phone width

Make the page work at 320 pixels wide with no sideways scrolling. Stack content on narrow screens, and keep buttons at least 24 by 24 pixels with space between them.
Keep everything else unchanged. Explain what you changed and how I'd check it.

Add a retrieval question at the end

At the end, add [one / three] low-stakes questions students answer from memory, with the earlier screens hidden. After each answer, show the correct answer and why.
Keep everything else unchanged. Explain what you changed and how I'd check it.

Signs: it undoes earlier fixes; it changes things you did not ask for; you have asked for the same fix twice and it is still broken; the chat is long and it has started to forget the spec; you hit a usage limit.

Then: update your spec with what you learned (new success checks, new red lines, new non-goals). Open a fresh chat, in the same tool or another, paste the updated spec and your latest working file, and ask it to continue from the file. Resume from your saved file and spec, never from the chat.

In Canva, paste the updated spec into a fresh chat, and check the new build against the version you saved.

"Resume from repository state, never chat history."

The rule in Alex Gray's own build, from Why a Normal Spec Isn't Enough When AI Builds the Thing

• Every success check, acceptance criterion and test case passes.
• The card 08 items for your level are ticked.
• Anything else goes in the spec as a version 2 idea.

"Detail is a cost. Spend it where a wrong guess is expensive."

Alex Gray, Why a Normal Spec Isn't Enough When AI Builds the Thing