Why Your AI Image Prompt Is Not Working: A Visual Debugging Decision Tree

The central idea

Prompt repair starts by naming the failure, not adding adjectives. Separate omission, contradiction, priority, spatial, counting, typography, and capability problems; each category requires a different next test.

A repeatable workflow

1. Name the visible failure

   Describe what is wrong in the output without restating the intended prompt or blaming the entire model.

2. Find the category

   Classify the issue as missing, conflicting, ambiguous, relational, textual, anatomical, or outside capability.

3. Simplify the test

   Remove unrelated instructions and rewrite the failed relationship in direct observable language.

4. Change the workflow

   When repeated tests fail, move typography, exact identity, factual diagrams, or precision work to a suitable tool.

Worked example

If “three glass bottles behind a ceramic cup” produces the wrong count and order, remove stylistic decoration, state the cup as the foreground focal object, place three separated bottles in a background row, and test the relation. If exact count remains critical, composite or photograph the arrangement instead.

Review checklist

• The failure is visibly described
• One root category is selected
• The repair prompt is simpler
• A workflow exit condition exists

Limitations

• Some counting and spatial failures are model limitations
• Repeated generation is not a substitute for a precision workflow