Edit only what I name below. Keep all other characters, including punctuation, spaces, and line breaks, except punctuation or spacing that belongs only to a deleted word. Without straightening, without fixing the tone, without shortening.
Change: [what to change, and how]
Return the full text, then list all the changes you made. If the same change is repeated, record it once with a count. If I specify all occurrences, I apply the edit to each match. If the goal or scope is ambiguous, ask before editing.
[PASTE YOUR TEXT]
I tested these prompts with a variety of more powerful “thinking” models, including ChatGPT (standard chat mode) on “high” reasoning, Claude Opus 5.5, and Gemini Argon.
For my message testing, I used a synthetic transcript of a 1,500-word voice note full of “ums,” “ahs,” and “you know” (“Okay, um, voice note of the week, it’s Sunday, uh, November 8”), and asked the models to “remove filler words from this transcript” and “return the full text.” When the models asked what were considered “filler words,” I told them to “use your own judgment.”
In my initial control runs without the “surgical edit” prompt, models were much more likely to make “collateral” edits throughout the document, formalizing some language (for example, turning “I’m going to” into “going to”), erasing barriers (such as “like okay” and “honestly”), and rephrasing casual phrases (turning “frame it as” into “frame it as”).
The “high” version of ChatGPT made between 25 and 35 of these collateral edits on my test transcript when I had not first used the “surgical” message, while Gemini Argon went from a minimum of seven unwanted edits to a maximum of 51. Claude Opus 5.5 was more careful without the guidance of the “surgical message”, only making between six and a dozen unauthorized edits.
