Once the model gives you an output that's almost good, the temptation is to say "rewrite it better". Bad move. Every vague word ("clearer", "more professional", "more concise") is a license to rewrite the whole thing, including the parts that were already fine.
The surgical correction prompt works the other way: you name the changes and you name what's not to be touched.
Same response but with these changes:
1. <specific change>
2. <specific change>
Don't touch the rest.Three lines. Numbered list so the model doesn't get lost. Explicit "don't touch the rest" close to prevent the silent rewrite.
Two named changes. A protective closing line. The output comes back with those two changes and nothing else.
This technique also helps when you pass code to the model and want it to fix ONE bug without touching the rest. "Change only line 12 so it returns null when the input is empty. Don't reformat the rest." If you don't say so, you get back the whole file with five cosmetic changes you didn't ask for.
On the right, the same follow-up in two shapes. Which one preserves the rest of the output?