When you write for somebody else's byline, “it doesn't sound right” is a problem you have to solve. Admiring the client's earlier books will only get you so far, because the next chapter has to reproduce enough of their choices to belong beside them. You need to work out what those choices are before you can judge whether your own pages make sense in that company.
Ghostwriting has taught me to examine voice in those terms. Once I can identify what a passage repeatedly does, I have something more useful than a vague feeling to guide the revision. That becomes especially valuable with AI, which can give me cleaner sentences while changing the narrator's habits. Without a clear idea of what changed, I can spend a long time arguing with a paragraph that fulfils my instructions and misses my intention.
I look beyond the vocabulary.
A narrator's voice begins partly in what she considers worth noticing. If she registers the practical problem in a room before she registers the people, her account will differ from that of a woman who first notices who has been excluded from the conversation. Their habits determine which details reach the reader and what those details seem to mean.
The sentences shape that account further. A dry qualification at the end of a long sentence lets the narrator revise what she has just told us, and the pleasure may depend on the reader following her thought far enough to need the qualification. Break the sentence into brisk statements and the vocabulary can remain intact while the timing disappears.
Emotional distance can change as easily. A narrator who usually leaves us to understand an evasive answer begins to feel different when every evasion comes with an explanation. Nothing in the revised passage has to be grammatically wrong for the reader to feel that someone else has taken over.
That is why a list of favourite words gives me only part of what I need. I also have to listen for how the narrator guides the reader towards an understanding, including the moments when she trusts us to get there ourselves.
A small change can alter the narrator's relationship with the reader.
Consider a hypothetical coastal saga whose chapters often end with a physical detail. After an argument, a woman puts out two cups, remembers that her sister has left, and returns one to the cupboard. An AI revision adds a concluding sentence to explain her difficulty accepting the separation.
The explanation may be accurate, although the action has already given the reader a chance to recognise that difficulty. If the narrator habitually leaves such moments open to interpretation, the addition changes her relationship with us. Repeat it often enough and the book begins to explain feelings we were previously allowed to discover. The sister has barely left the kitchen, and someone is already explaining the cups.
I would compare the passage with neighbouring scenes before deciding whether to remove the sentence. Some books depend on a narrator who interprets events openly, so I want to see how this one handles similar moments. That comparison gives me a reason to accept or reject the addition that a general instruction to make the prose more emotional cannot provide.
The brief gives me a way to judge the change.
Writing for an established readership also makes me consider what happens when I alter a familiar convention. The assignment concerns a particular kind of book, and I need to understand what its readers expect well enough to recognise when a revision changes the experience.
A Gothic heroine who sensibly refuses to enter a dangerous house illustrates the difficulty. Her refusal may fit the character perfectly, but it also leaves me to find another way for the danger to reach her. Before I congratulate her on her improved judgement, I have to work out what happens to the story once she goes home.
AI suggestions deserve the same examination. Making a character more agreeable can remove the friction a scene needs, while cutting a repeated interaction can remove a deliberate escalation. Either change might improve the manuscript; I need to understand what the original passage contributed before I can tell.
I ask for comparisons before I ask for repairs.
When a passage sounds wrong, I give the model a sample that already works and ask it to compare a specific feature. It might examine how much emotional explanation each passage contains, with quotations that let me check whether it has identified a real difference. An assertion that one version is more immersive offers me far less than a clear account of where the revised narrator begins to explain an emotion previously conveyed through action.
With that difference in view, I can decide whether I want to repair it. Removing one redundant explanation while preserving the paragraph gives me a change I can inspect. Restoring the emphasis of a sentence gives me another. Either request is easier to assess than a full chapter rewritten under the instruction to sound more like me.
It also leaves room for the revisions I do want. I can accept a clearer sentence without accepting every change that accompanied it, because I have a reason for keeping one and rejecting another. The useful question becomes what this narrator would say here and how she would expect the reader to understand it.
Ghostwriting has made me more attentive to those choices in my own prose as well as in the voices I write professionally. When an editing tool keeps trying to remove a habit I want to protect, I can point to what that habit does for the book. I would rather have that disagreement than discover, several chapters later, that I have accepted a very competent stranger as my narrator.