How to fix AI-written dialogue where every character sounds the same

When every character in an AI-written chapter speaks in the same tidy, polite sentences, you can fix it in Write Studio. Give each speaker a short voice note in the Story Bible, select the conversation, choose Rewrite, and then check the result with one simple test: hide the speaker tags and see whether you can still tell who is talking. This walkthrough takes one six-line exchange from a flat draft to a saved revision, without changing what anyone says.

Before you start

Step 1: Hide the speaker tags and read the exchange

A speaker tag is the “she said” or “Theo said” that tells you who is talking. Cover or delete the tags in a copy of the conversation and read the lines alone. If you cannot tell who says which line from the words, the characters are sharing one voice. In an AI draft this can happen when nothing tells the model how each person talks.

In this example, Nell is a retired ferry mechanic and Theo is her nephew. Remove the tags and every line could belong to either of them: similar length and the same polite structure. Four of the six lines open with “I”, and three turn on “But”. Before changing anything, write down what the conversation has to keep. These six points are your check later, because a revision that sounds better but drops or adds a point has changed the scene.[5]

Step 2: Write a short voice note for each speaker

A voice note tells the AI how a person talks, not what they say in this scene. Keep it to a few concrete habits and two sample lines. Useful habits are the ones you can hear: how long their sentences run, the words they reach for, what they never say, and what they do when they are pushed. Write sample lines that are not from the scene you are fixing, so the AI learns the voice rather than pasting the lines in. Another tool’s documentation recommends a similar approach: give the AI examples of how a character speaks, because brief summaries tend to produce stereotypes.[4]

Make the two voices contrast on more than one habit. Sentence length is the easiest difference to hear, so it is a good place to start. Avoid giving a character a catchphrase to repeat in every line; one or two habits per line is enough to identify a speaker.

Step 3: Add the voice notes to the Story Bible

The Story Bible is the pane on the left of Write Studio. It holds narration and voice, characters, setting and continuity facts, and every AI action reads it, up to 16,000 characters in total. That is why the voice note belongs there rather than only in a one-off instruction: once it is in the Bible, later rewrites and continued scenes read it too.[1]

If you have not generated your book yet, you can do this up front instead. The Pro and Ultra Max generators accept your characters, places and rules before generation, and those are seeded into the story bible before chapter one is written. If you include each main character’s voice note there, it is in the story bible from the start, and you can still check and adjust it in the Studio afterwards.[2]

Step 4: Select the conversation and choose Rewrite

In the manuscript editor in the middle, highlight the conversation, including its tags and the short actions between lines, but not the paragraphs around it. A smaller selection keeps the AI focused on the voices. Then choose Rewrite. This starts an AI action immediately, so add the voice notes before you choose it. The suggestion streams in as tracked changes, with new words in green and removed words in red. Wait for it to finish, and do not choose Accept yet.[1]

Step 5: Check the suggestion with the tags hidden

When the suggestion has finished, read all of it, not just its first lines, and run these checks before you accept or steer anything:

Your suggestion will use different words. The target shows what passing looks like. With the tags hidden, Theo’s lines are the longer, qualified ones that use his habits (“honestly”, “I just think”, “Aunt Nell” when he pushes), and Nell’s are flat statements under eight words. Each line keeps its point from the table: Theo calls selling sensible; Nell’s “Fair enough” acknowledges why he thinks so and “She matters, though” keeps the boat’s importance without explaining her feelings; Theo grants that and gives the repair cost; Nell says maybe but not yet; Theo asks her to at least think about it; Nell agrees to that and no more. Theo’s habit words come from the list in his note, but no sample line or phrase from one (such as “in fairness to everyone”) appears, and no new reason is added. In line 3, “the repairs” stays general; a version that named the hull, quoted a price, or added a reason such as “for me” or “for everyone” would add something the scene never had.[5]

Step 6: If it fails, steer it with a note and choose Apply

If the suggestion passes every check in Step 5, skip to Step 7. If the voices are still too close, or a line fails another check, type a note in the instruction box beneath the preview and choose Apply. Change the names to match your book.[1]

Apply generates a new suggestion for the original selection using your note. It does not accept anything. Each successful generation spends another action: a Rewrite plus one Apply is two actions, 1,000 credits at Fast or 6,000 at Pro if both return text. Try again repeats the last applied note, so if you type a new note, choose Apply rather than Try again. When the new suggestion finishes, run the Step 5 checks on it again.[1]

Step 7: Keep or reject it, then check it saved

If it still goes wrong

For the rest of the book

Because every AI action reads the Story Bible, Continue the scene and later rewrites read the voice notes too. Run the tag test on new conversations as you go, and update a character’s note if their voice should change over the story. If you are on a Pro plan, the Deep Manuscript Audit lists Dialogue among its twelve categories and quotes each passage it flags, at 40,000 credits per chapter. Use the same test on anything it raises.[1][2]

Limits

A voice note is guidance the AI reads, not a guarantee, so review every suggestion before you accept it. Voices only go so far: if two characters want the same thing in a scene, the conversation may need a different disagreement, not different wording. The tag test and voice notes are an editorial method we have not measured, and the before, the voice notes and the target above are invented examples rather than recorded AI output.