Prompt Tests Show How to Stop ChatGPT and Claude From Over-Editing Text

AI chatbots like OpenAI's ChatGPT (running GPT-6.1-Sol), Anthropic's Claude Opus 5.5, and Google's Gemini Argon frequently go beyond user requests by altering prose, formalizing casual words, and stripping out natural conversational phrasing. To address this over-editing tendency, recent testing evaluated a custom 'surgical edit' prompt designed to limit AI models strictly to targeted revisions while preserving original sentence structure and voice.

During benchmark tests on a synthetic 1,500-word voice-memo transcript filled with filler words, standard editing requests without strict prompt constraints resulted in widespread unauthorized modifications. High-reasoning ChatGPT made 25 to 35 unwanted collateral edits, while Gemini Argon executed between 7 and 51 unsanctioned changes, such as converting 'gonna' into 'going to' and removing conversational hedges. Claude Opus 5.5 proved most restrained in baseline tests, committing only 6 to 12 collateral edits.

Implementing the dedicated surgical prompt reduced unwanted modifications to near zero across tested platforms. Claude Opus 5.5 and Gemini Argon eliminated unauthorized cuts entirely, while ChatGPT retained only a minimal number of questionable edits. A longer version of the prompt, which requires the AI to itemize every change made, also caused the models to actively ask for clarification on ambiguous instructions in five out of six runs.

Should you use it? For simple single-word adjustments, editing manually is still faster. However, for users processing lengthy transcripts or long-form prose who need to strip filler without sacrificing their personal writing style, employing a surgical prompt is an effective way to keep AI editing in check.