Transcribe two minutes of a real conversation, word for word, and read it back. It is almost unusable. People talk over each other, abandon sentences halfway, repeat themselves, say "like" and "you know" constantly, spend forty seconds establishing that yes, it did rain earlier. Anyone who has looked at transcripts of natural speech knows this: real conversation is mostly repair and filler, and it works in person because tone, face and shared context are carrying most of the load.
So the goal is not realism. It is the impression of realism, which is a constructed thing, and which requires keeping a few specific features of real speech and discarding almost everything else.
Keep: interruption, incompletion, evasion
The features worth preserving are the ones that carry information about the relationship.
Interruption tells you who has power and who is desperate. A character who is cut off mid-sentence, repeatedly, by the same person, has been characterised without a word of description.
Unfinished sentences tell you where feeling exceeds language. "I just — no. Forget it." does more than any completed version of that thought.
Not answering the question is the most useful of the three, because it is how people actually deflect. Ask someone whether they are angry and they will tell you the restaurant is booked for eight. The reader understands immediately.
Cut: filler, pleasantry, and information the reader does not need
Everything else in a transcript goes. The hellos, the weather, the ums. Fictional characters can arrive mid-conversation and leave before it ends, and readers assemble the missing parts without noticing. Start every exchange as late as possible and leave as early as you can bear.
The hardest thing to cut is information the writer wants delivered. Characters explaining to each other things they both already know — the shape of their shared history, how the technology works, what happened last Tuesday — is the most reliable killer of dialogue, and it is obvious from outside and invisible from inside. If someone says "as you know", or if a character asks a question purely so another can answer it, the scene has stopped being a conversation and started being a delivery mechanism.
The whole craft is subtext
Good dialogue has two layers: the thing being discussed, and the thing actually at stake. Whenever those two are the same, the scene goes flat.
Consider a couple arguing about a suitcase. One of them is leaving for a job in another city.
"Don't go." She was looking at the suitcase, not at him.
"Sare—"
"No. I'm not doing the long-distance thing. I can't."
He sat down on the case, which deflated slightly under him. "You don't get to decide this on your own."
"Neither do you." She finally looked up. "So we're stuck."
Nobody in that exchange states the actual question, which is whether either of them will give up their life for the other. They talk about logistics and about who decides. The reader supplies the rest, and the reader's supplying of it is what makes the scene work — a character who explained the stakes out loud would have removed the reader's job.
The general rule: characters should want something from each other in every conversation, they should mostly not say what it is, and what they say instead should be a strategy for getting it.
Give each person a different mouth
The test is mechanical. Delete all the attributions from a page of dialogue, hand it to someone, and ask them to tell you who is speaking. If they cannot, you have written one character having a conversation with themselves.
Difference does not come from accents or verbal tics, which are the first things writers reach for and the least effective. Phonetic dialect on the page is difficult to read and ages badly. What actually differentiates people is structure: sentence length, how much they hedge, whether they answer directly, whether they ask questions, how much they are willing to leave unsaid.
An anxious person qualifies everything and asks for reassurance inside the sentence. A person used to being obeyed uses short declaratives and does not soften them. An academic reaches for precision when precision is not required. Someone speaking a second language may be more formal, not less fluent — textbook grammar in a context where a native speaker would be slangy.
Here is the same difference doing character work in four lines:
"That's not what Kant argued. The Critique explicitly—"
"I know what Kant argued. I read it." She kept stirring her coffee. "I disagreed with your reading of it."
He stopped. "Well — most people don't get that far in."
"I'm not most people."
No one has been described. One of them needs to be the person who has read the book; the other has read it and does not need anyone to know.
Attribution: said, and the argument about it
Elmore Leonard made this the centrepiece of his rules of writing: never use a verb other than "said" to carry dialogue, and never modify "said" with an adverb. The reasoning is that "said" is invisible — readers process it without registering it — while "he expostulated" and "she said angrily" both put the author between the reader and the scene.
This is good advice and it is worth knowing that it is a stylistic position rather than a law. Leonard was arguing for a particular kind of transparent, hard-boiled prose, and plenty of excellent novelists use a wider range of attribution perfectly well. What is not in dispute is the underlying diagnosis: an adverb attached to "said" is almost always doing a job the dialogue itself failed to do.
"I'll do it," he said determinedly, is a writer telling you how to hear a line that should have been written so you could hear it.
The better instrument is action. Instead of an adverb, give the character something to do:
"I'll do it."
She raised an eyebrow. "Really."
He straightened up. "Of course."
The beats do three things at once — they attribute the line, they carry the emotional register, and they keep the scene physically located in a room. A page of unbroken dialogue starts to float free of any place; a beat every few exchanges anchors it.
Read it aloud, always
This is the only revision technique for dialogue that I would call non-negotiable, and it catches things the eye does not.
Your ear will find the sentence that no human mouth would produce. It will find the four consecutive lines with identical rhythm. It will find the place where a character says something that is true and well-phrased and completely out of register for them. And it will find the exchanges where both people are being equally articulate, which is the commonest failure of all — in real disagreements, one person is usually much better at talking than the other, and that asymmetry is where the tension lives.
Listening as research
Sit somewhere public and pay attention to how people actually talk, not to what they say. Notice how often someone answers a different question from the one asked. Notice how much is settled by tone. Notice that the most emotionally loaded conversations tend to use the plainest vocabulary — people in genuine distress do not become eloquent, they become monosyllabic.
Then write none of it down verbatim. The point of the exercise is calibration, not material. You are training an ear so that when you write a line, something in you knows whether a person would say it.
Dialogue that reads as effortless is the most heavily worked prose in a novel. Readers will not notice any of the craft, which is the entire objective — they should just hear two people in a room, wanting different things, neither of them saying so.
Tags
Taresh Sharan