Edit only what I name below. Preserve every other character, including punctuation, spaces, and line breaks, except punctuation or spacing that belonged only to a deleted word. No smoothing, no tone fixes, no shortening.
Change: [what to change, and how]
Return the full text, then list every change you made. If the same change repeats, list it once with a count. If I specify all occurrences, apply the edit to every match. If the target or scope is ambiguous, ask before editing.
[PASTE YOUR TEXT]
I tested these prompts with a range of more powerful “thinking” models, including ChatGPT (standard chat mode) on “high” reasoning, Claude Opus 5.5, and Gemini Argon.
For my trials with the prompt, I used a synthetic 1,500-word voice-memo transcript filled with “um’s,” “ah’s,” and “you know’s” (“Okay, um, voice memo for the week, it’s Sunday, uh, November 8th“), and asked the models to “remove the filler words from this transcript” and to “return the full text.” When the models asked what counted as “filler words,” I said to “use your own judgement.”
In my initial control runs without the “surgical edit” prompt, the models were far more likely to make “collateral” edits throughout the document, formalizing some language (for instance, turning “gonna” into “going to”), erasing hedges (like “kind of agree” and “honestly”), and rewording casual phrases (turning “frame it like” into “frame it as”).
The “high” version of ChatGPT made anywhere from 25 to 35 of these collateral edits in my test transcript when I hadn’t first used the “surgical” prompt, while Gemini Argon went from a low of seven unwanted edits to a high of 51. Claude Opus 5.5 was more careful without the “surgical prompt” guidance, only committing between six and a dozen unauthorized edits.