Editing Dialogue and Quotations with AI

Signs of Dialogue That Rings False

False dialogue comes from one habit: writing conversation like a report instead of like behavior. The tells are countable.

• Everyone speaks in complete, well-formed sentences. Real people cut in and trail off.
• Everyone sounds the same. Take away the speech tag and you cannot tell who is talking.
• Characters explain to each other what both already know, purely so the reader catches up.
• Every question gets a direct answer immediately, even when answering costs the speaker.
• Tone is carried by the adverb rather than the words: "he said angrily," "she said hesitantly."

"So you are the one who moved the file."

"Yes, I moved the file, because I wanted to protect my brother."

Both lines hand over the information and hide nothing. A better version delivers the same information through resistance and delay; that small effort is what makes dialogue live.

What AI Sees and What It Cannot See

A language model is good at patterns: the repeated "he said," the flat rhythm of exchange after exchange, inconsistent punctuation, and sentences shaped identically for every character. It will find these across hundreds of pages.

What it cannot see is subtext: who is lying, what is being hidden, what a line gestures at without naming. The model does not know what you deliberately kept out of a character's mouth.

A practical warning: a model can build a plausible-but-wrong quotation, a line with confident attribution that appears nowhere in your sources. Any attribution you did not write yourself is suspect until you check it.

The character sheet you hand over

Write a few lines per character: age, work, what they want, what they hide, and how they talk (short sentences, talkative, ironic). The model uses that sheet to keep speech patterns consistent; it does not build subtext from it. Keeping your voice intact through the edit is separate work, covered in keeping your voice while editing with AI.

Standardizing Your Quotation Style

Choose the rule first, then let the model hunt for deviations. The rule has four parts: the opening and closing marks, quotations inside quotations, where terminal punctuation sits, and how paragraphs break when the speaker changes.

The decision is yours, because the model does not know which convention you want. The sweeping is the model's job, because it sees the whole manuscript at once and will not miss an instance.

The quotation style for this book is at the top of the file. Search the whole manuscript and list only the places that deviate from it. Do not rewrite anything. For each case, give the location and the deviating form.

The output is a list you accept or reject one line at a time. It needs the same accept-and-reject discipline described in the line editing checklist; the only difference is that the unit of work here is a quotation rather than a sentence.

Attribution: Where Editing Stops and Fact-Checking Starts

Editing may clean up the form of an attribution: where the speech verb sits, the punctuation, the repetition of "said." Editing must not invent a quotation or make one tidier.

Two rules. A quotation you cannot verify stays untouched or gets cut. And never ask a model to fill in the gap in a partial quotation. Nevispar has no automatic source verification and no web search, so checking that a quotation is real and attributed to the right person is your work, and it stays your work.

The point that gets lost most often: the broken edges of real speech are data, not errors — dialect, an unfinished sentence, a repeated word. A model that "fixes" them corrupts the record and flattens the voice.

When attribution appears inside narration, you are back at the line-editing boundary: standardize "said" and "asked," but never put words in someone's mouth. If you are unsure where these levels belong, the complete map of smart editing sets the order.

Direct and Indirect Quotation

Choosing between the two is an editing decision, not a preference. The table below lays out the difference.

The working rule: quote directly wherever a claim, a figure, or eyewitness testimony is at stake; report indirectly when you are orienting the reader or summarizing. One short direct quotation carries more weight than a paragraph of paraphrase and takes up less room.

A Pre-Delivery Checklist

Run this on the final version, not the draft; the last few items overlap with sentence-level work, but here they ask only about talk and quotation.

• Every direct quotation has been checked against the original source.
• Every attribution carries the right name and the right role.
• No quotation is left half-finished or trailing a pointless ellipsis.
• Punctuation style is identical throughout the book.
• Paragraph breaks line up with changes of speaker.
• Long exchanges have somewhere for movement or silence to happen.
• No character explains what both parties already know.
• Tone comes from the words, not from "he said angrily."
• The characters' voices are still distinct after editing.
• Every suggestion you accepted can be explained.

Dialogue and quotation are the most fragile part of a manuscript under automated editing: voice and subtext on one side, the record and attribution on the other. A model can produce a list, flag an inconsistency, and propose an alternative to a repeated line; the final call, and responsibility for every quotation being right, stay with you.