What to do when an AI project fails: a 20-minute review
https://patrickz.ai/learn/learn-from-a-failed-prototype/

OUTCOME
Write a failure review that produces one smaller, testable next attempt.

YOU NEED
A disappointing AI result and the prompt that produced it. No failure handy? Ask any chat or image tool for “a poster for our bake sale” with no other details, and review that.

1. Name the intended result
Write what you expected to happen and the evidence that it did not. Keep observation separate from your explanation of the cause.

2. Classify the gap
Ask whether the problem lies in the brief, input, tool capability, implementation or evaluation. More than one cause may be possible; list uncertainty.

3. Learn one missing fundamental
A fundamental is a basic principle of the craft. Name the smallest one that would have prevented the problem, in plain words: text you can read from across a room, information that survives closing the page, a source you can trust, or the facts the tool needed. Ask for one short exercise on just that principle.

4. Repeat a smaller test
Change one thing and rerun. Keep the failed version and the result. Decide what evidence would justify moving back to the larger project.

PROMPT
Help me review this failed attempt without generic encouragement. Intended result: [goal]. Observed result: [evidence]. Original instructions: [prompt]. Separate facts from hypotheses, identify one missing fundamental, and propose a small next experiment with a pass/fail check.

EXAMPLE / EXPERIMENT
A bake-sale poster with no date, place or price has a brief problem: add those three facts, change nothing else and rerun. If the facts are there but nobody can read them from across the room, the missing fundamental is text size and contrast, and a newer image model will not fix it. A game character that moves faster on a faster computer needs a lesson on frame timing (often called delta time) before new artwork.

CHECK YOUR RESULT
[ ] The review describes an observable gap.
[ ] The next test changes one thing.
[ ] You can explain what you learned even if the new attempt fails.

IF IT FAILS
If the review becomes a long roadmap, reduce it to one uncertainty. Keep the practice project low stakes so experimentation stays reversible.

MY RESULT / NEXT CHANGE