How to Write Believable Dialogue: The Subtext Rule

Read your dialogue out loud. If it sounds like two people efficiently exchanging accurate information, you've written a transcript, not a scene. Real people are terrible at saying what they mean, and that incompetence is the entire reason dialogue is interesting.

Most advice on how to write dialogue gets stuck on mechanics: dialogue tags, when to use "said," how to format. Useful, but it's the woodworking, not the design. The thing that actually separates wooden dialogue from dialogue that crackles is one principle, and almost everything else flows from it.

The subtext rule: people don't say what they mean

Here is the rule. In a good scene, the words on the surface and the meaning underneath are not the same thing. The gap between them is the subtext, and that gap is where the reader leans in.

A couple is arguing about who left the dishes in the sink. The dishes are not the point. The point is "I feel like I do everything and you don't see me." Neither of them will say that sentence, because people almost never say the real sentence. They fight about the dishes. The reader, watching the dishes-fight, understands the marriage-fight underneath, and that understanding is the pleasure.

When you write a line of dialogue, ask: what does this character actually want, and are they saying it directly? If they're saying it directly, you've probably written it wrong. Find the thing they'd say instead, the deflection, the smaller complaint, the deliberately wrong question. That's the believable line.

Watch the difference.

On the nose: "I'm angry that you never call me anymore. It makes me feel abandoned, like our friendship doesn't matter to you."

With subtext: "Oh, you still have my number. Good to know."

The second line carries the same wound and never names it. The reader does the work, and readers love doing the work.

Every exchange needs conflict

This is the second rule and it's not optional. Every scene of dialogue needs friction, even between people who like each other. Not necessarily an argument. Friction can be two people wanting slightly different things from the same conversation.

One wants to confess, the other wants to keep things light. One is trying to leave, the other is trying to make them stay. One wants the truth, the other wants comfort. The moment two characters are in perfect agreement and after the same outcome, the scene goes slack, because there's nothing for the dialogue to do.

A simple diagnostic: for any conversation, write down what each character wants from the other in that specific moment. If both answers are the same, or one of them is "nothing," cut the scene or find the buried disagreement. Dialogue without competing wants is just two people agreeing in public, which is pleasant in life and death on the page.

Cut the hello, cut the goodbye

Almost every first draft opens conversations too early and closes them too late. The greetings, the small talk, the "how was the drive," the wind-down where everyone says bye, none of it earns its place.

Start the scene at the moment of friction and end it on the line that lands hardest. If two characters meet for coffee to discuss a betrayal, you do not need:

"Hey, thanks for coming. Did you find parking okay? I'll have a latte. So, how have you been?"

You need to start three lines into the real conversation, the moment one of them says the thing the scene is actually about. Trust the reader to assume the greetings happened offscreen. Nobody has ever closed a book because a character didn't say hello.

Same at the end. The instant the scene has delivered its turn, get out. Don't let your characters politely conclude. The cut is part of the punch.

Distinct voices, or it all sounds like you

If you covered the dialogue tags, could a reader still tell who's speaking? For most drafts the answer is no, because every character talks in the author's voice: same rhythm, same vocabulary, same level of wit.

Real distinctiveness comes from how people talk, not catchphrases. Sentence length (does this character speak in fragments or run-ons?). Vocabulary register (does she say "purchase" or "buy"?). What they evade versus what they hit head-on. The nervous one over-explains. The powerful one uses short sentences and lets silence do the threatening. The one with something to hide answers questions with questions.

The drill is simple and merciless: take a page of dialogue, delete every tag and name, and hand it to someone. If they can't reliably say who's speaking, your characters don't have voices yet, they have your voice in different hats. Getting voices distinct is the same muscle as keeping your own voice intact while drafting: it lives in the specifics of how a particular mind meets the world.

The on-the-nose trap

"On the nose" is the term for dialogue that states exactly what's happening or what a character feels, with no gap, no subtext, nothing for the reader to infer. It's the single most common dialogue problem and it shows up worst in two places.

The first is emotional confession. "I'm scared." "I've always loved you." "I feel so guilty about what I did." Occasionally a character should say one of these flat out, and when they do after a whole book of not being able to, it's devastating. But used casually, it robs every line of its power. Make characters circle the feeling, not announce it.

The second is exposition. "As you know, Captain, our ship has been adrift since the reactor failed three days ago." Nobody tells someone what they both already know. If two characters share information, only one of them should be learning it, and they should be learning it because they want something. We dug into this trap, and the wider fix, in the piece on worldbuilding without the info-dump.

Using AI without flattening the voice

AI is genuinely dangerous for dialogue, because the default output of any language model is competent, fluent, and completely on the nose. Ask a model to "write a tense conversation" and it will hand you characters who articulate their feelings with perfect clarity, which is the opposite of tense.

So don't generate dialogue, diagnose it. Paste your scene and ask whether any line is too on the nose, where the subtext is doing real work and where it's missing, and whether the two voices are distinguishable with the tags removed. That's the job of an AI writing coach: it doesn't write the lines, it tells you which of yours are lying flat. If you've been using a generator and the dialogue keeps coming out smooth and dead, this is the argument for a Sudowrite alternative built to sharpen your writing instead of replace it.

Read it aloud, then cut a third

The last technique is the oldest one and the one writers skip most. Read every line of dialogue out loud. Your ear catches what your eye forgives: the clunky rhythm, the line no human would say, the speech that runs three sentences past its point.

Then cut. Most first-draft dialogue is at least a third too long. Trim the throat-clearing, trust the subtext, and let the silences carry weight. The lines that survive will sound like people, which is the only thing dialogue ever has to do.

---

Polyz is a manuscript editor with AI tools and an AI writing coach that helps you find on-the-nose lines and sharpen your dialogue, without writing it for you. Start a free trial.