Key answer
Natural-language punctuation need not be perfect, but that does not mean punctuation never affects output. Distinguish changes in meaning from sensitivity to formatting when meaning stays the same.
Full explanation
“No, continue editing” and “do not continue editing” are opposites. Unclosed quotation marks can obscure the boundary between source text and instructions. Correct these because meaning or boundaries have changed.
*When Punctuation Matters* compared formatting changes in Llama, Qwen and Gemma models, with additional GPT-4.1 and DeepSeek V3 tests. Formatting alone caused differences of approximately 8–10 accuracy percentage points on some tasks. These were specific classification, multiple-choice and math settings, not evidence that standardizing everyday Chinese punctuation improves accuracy by 8–10 points. Section 4.4 and limitations
Prioritize punctuation affecting interpretation: negation scope, grouped conditions, decimal points and material/instruction boundaries. A missing final period rarely justifies rewriting a whole prompt. For recurring fixed tasks, compare several clear formats and choose a more stable version empirically.
JSON, code and other machine formats are different: quotation marks, commas and brackets may be syntax. Errors can make parsing fail outright and cannot be treated like tolerable natural-language imperfections.