Does asking an AI to fix just the sentence you've highlighted work better than asking it to fix the whole document at once?

This asks whether prompting an AI about a specific highlighted passage gets better results than one prompt covering the whole document. The short answer: the collection has no head-to-head study of selection-scoped versus global prompts, so it can't settle that directly. What it does offer is several findings that explain why scoping a prompt might matter, and why the answer probably depends on the task more than on a general rule.

The strongest theoretical case for highlighted-selection prompts comes from looking at what a prompt actually is. In human conversation, people build shared context gradually and adjust it as they go. A prompt squeezes all of that into one fixed frame that the model can't renegotiate, so changing direction mid-task means writing a new prompt instead of making a quiet adjustment How do prompts reshape the role of context in AI conversation?. Read that way, a highlighted selection does part of the context-setting for you: the selection itself says "this part, not the rest," which a global prompt has to spell out in words. Related work on chain-of-thought prompting supports this. Reasoning prompts fail when the question's meaning doesn't reach the prompt before the model starts working Why do some questions perform better without step-by-step reasoning?. A tight, clearly bounded target makes that easier.

There's also evidence that where material sits inside a prompt shifts results more than you'd expect. Moving an identical block of examples from the start of a prompt to the end can swing accuracy by up to 20% and flip almost half the answers How much does demo position alone affect in-context learning accuracy?. Compression research shows that instructions and questions carry far more weight than surrounding material, which can be trimmed heavily with little loss Can prompts be compressed without losing model performance?. Together these suggest that a global prompt, where the passage you care about is buried somewhere in a long context, could be at a real disadvantage compared with a prompt that puts the selected text up front. This is an inference from adjacent findings, not a tested result.

The counterweight is that prompt techniques rarely work the same everywhere. A benchmark of 23 prompting styles across 12 models found that methods which help cheaper models can hurt stronger ones, and that the structure of the task decides what works Do prompt techniques work the same across all LLM tiers?. A selection prompt that cuts off the surrounding document may lose the context needed for tone, consistency, or references elsewhere in the text. That's a plausible case where a global prompt would win.

The angle you might not expect is a social one. In a study of paired writers using shared AI-assisted editors, people strongly preferred seeing when, how, and *where* their collaborator had used AI, because it helped them understand the other person's thinking and check AI-generated text Do writers want to see each other's AI prompts in shared editors?. Selection-anchored prompts tie each AI action to a specific stretch of text, which makes that kind of visibility possible. Global prompts can't do that as easily. So even without accuracy data, "contextual vs. global" may matter as much for how people collaborate and stay accountable as for output quality.