Skip to content

Which Passages the Reviewer Chose

When AI revisions depend on a reviewer's selected passages, the handoff must preserve that selection and keep the reviewer's instructions separate from the model's interpretation.

A Reddit discussion about Word highlights and comments records a document-revision handoff. The original poster describes using ChatGPT to extract highlighted text and associated comments into a Markdown text file, then giving that file and the original document to Claude for revisions. Other commenters report different experiences, including successful comment reading. These are individual accounts, not a controlled application comparison or a verified remedy.

For anyone delegating revisions, the distinction matters. An assistant can identify sentences it considers important without identifying the sentences a reviewer selected. It can also find a selected passage without knowing whether to keep it, question it or change it. Useful revision work needs the selection and the instruction; the model’s judgment cannot quietly stand in for either.

An October 21 preprint compares representations of 76 selected policy documents. A model retrieving marked requirements scored 0.690 with the researchers’ annotations and 0.075 with plain text. The measure, F1, balances misses against wrongly included items; these are not document-accuracy percentages. Formatting supplied the answer key, without human validation. This measures mark recovery, not compliance or verified author intent.

Which words were selected?

Consider two hypothetical copies of the same draft. The sentences are identical, but a reviewer has highlighted the introduction in one and the conclusion in the other. Convert both to identical plain text and the difference disappears. The assistant can still understand both passages. It cannot tell from that text which one the reviewer chose.

This is a different failure from misreading a sentence. The wording has survived, and may even be quoted perfectly. What is missing is a relationship between a person and a passage: this is where they directed attention. A more persuasive summary of the document cannot reconstruct an arbitrary selection that left no surviving trace.

Word files need not force that loss. Text and its presentation are distinguishable parts of their structure. The python-docx documentation describes paragraphs containing smaller pieces called runs, each with text and formatting attributes. A converter can keep the words while omitting attributes attached to them. Readable output therefore establishes something useful, but narrower than a faithful record of the review.

Nor is preserving the attributes an unsolved impossibility. The docx2python project’s documentation describes an option that exports basic formatting as HTML, including colored text and highlights. That is documented capability, not a test of a particular chat application’s upload path. It makes the engineering question concrete: which distinctions does the chosen route retain, and which does the next task need?

That does not make every visual property essential. A question about an argument may depend only on its wording. A question about a reviewer’s selection depends on the marks. The appropriate handoff follows the work being requested, rather than treating either plain text or full visual preservation as the right answer for every document.

A mark can select without instructing

In December 2024, a user asked for something quite specific in the OpenAI Developer Community: analyze a Word document while taking account of passages they had marked with highlights and font colors. The dated feature request does not establish what today’s applications can read.

The request joins two operations that are worth keeping distinct: recognizing a mark and deciding what to do because of it. The first can establish that a passage was colored red. It cannot, by itself, establish that red means the passage is correct, urgent or approved.

Imagine a review in which red means “check this claim.” Recasting those passages as the most authoritative statements would reverse the instruction. The color would have survived, but the assistant would be acting on an invented explanation of it. A document legend or the reviewer’s own directions can supply that missing meaning. A color alone cannot.

Nor does preserving a highlight require turning it into a command. Someone may mark a passage for discussion before deciding whether it should change. An assistant can report the selected text faithfully and leave the requested action unresolved. If it separately recommends a revision, the recommendation should remain its own, not become something the reviewer supposedly asked for.

Keep the connection through the edit

The opening workflow carries an original document alongside an extracted account of its review. Its useful object is not just a list of sentences. A comment needs an identifiable passage; a highlight needs an identifiable span of text. Put the comments and passages into unrelated lists and the next reader has to reconstruct the connections, even if no words are missing.

Keeping those connections is a design aim, not a demonstrated productivity gain. It also gets harder once the draft changes. A passage may move, be rewritten or disappear. The original selection can still document what the reviewer saw, but it does not automatically specify where the next edit belongs. Carrying the review forward includes preserving uncertainty about that match, not silently attaching a comment to whichever new sentence seems closest.

This leaves a concrete choice for an assistant preparing revisions: distinguish the passage the person selected, the instruction the person supplied and any new recommendation. A handoff that preserves those differences lets the next stage continue the review. A handoff that merges them asks the model to replace part of the reviewer’s work before the editing has begun.

Sources

Understand How People and Agents Work Together.

Practical essays, experiments, and lessons from building software for people and agents.

Occasional emails, sent when there is something worth reading. Free to subscribe. Unsubscribe at any time.