Reduce broken text in generated images

Plan the text in this generated image.

First, sort each piece of text:
1. **Generate it** — very short, Latin alphabet, not load-bearing. Even then expect several attempts.
2. **Leave a blank and composite later** — anything long, anything in Korean or another non-Latin script, anything that must be exactly correct: prices, dates, brand names, contact details. *A misspelled price is not a style problem, it is a wrong image.*
3. **Remove it** — text that is there out of habit and does nothing.

For group 1, write the prompt clause: the exact string in quotes, where it sits, and its relative size.
For group 2, specify the negative space: where, how large, what background it sits on so the overlay reads.

Then:
- The prohibition list — generators insert uninvited watermarks, captions, and logos. These go in every prompt whether or not you wanted text.
- How to check the output: read every character out loud. Near-miss letterforms are the common failure and the eye skips them.
- ⚠️ Text rendering is improving fast between model versions. *Note that this whole plan may be unnecessary within a generation or two* — re-test before assuming it still applies.