Dialogue reads fake for two mechanical reasons before any question of talent: every line is the same length doing the same informational job, and the attribution carries flourishes the reader trips over. Elmore Leonard's rule — never use a verb other than "said" to carry dialogue, from his rules of writing published in The New York Times in 2001 — remains the cleanest fix for the second problem. The first problem, sameness, is repaired by making every line an action a character performs on another character.
What is dialogue actually for in a scene?
It is characters trying to get things from each other. In a working scene, each line is a move: a probe, a deflection, a bid, a refusal. When dialogue sags, it is almost always because the lines have stopped being moves and become delivery — two voices distributing exposition the author needs placed. The test is brutal and simple: cover the dialogue and check whether anything changed between the characters. If their positions are identical at the end of the exchange, the conversation was scenery.
This is why transcription realism fails too. Real speech is full of greeting rituals, restarts, and weather; readers will not sit through it, and the complaint that dialogue "doesn't sound real" almost never means it should sound recorded. It means it should sound compressed the way memory compresses — the sense of a conversation without its dead air.
Why do all my characters sound the same?
Because they share one nervous system — the author's — and identical sentence rhythm across speakers is the tell. Voice on the page is built from grammar habits, not vocabulary lists: one character speaks in complete sentences and qualifies everything, another answers questions with questions, a third goes quiet and then says the one thing that matters. The diagnostic exercise is the names-covered test: strip the tags from a page of dialogue and ask a reader to assign lines to characters by rhythm alone. If they cannot, the voices are not yet built.
Repair comes before drafting new scenes, in a voice sheet per character: sentence length defaults, formality, what they refuse to name, their verbal tic that appears exactly twice per book. Two entries in the sheet is often enough to separate voices; the full personality questionnaire mostly produces homework that never touches the page.
How should attribution work?
Leonard's advice remains the industry default: "said" is invisible, and verbs like chortled, snapped, and admonished draw attention to the writing instead of the speech. Modifying adverbs — he said angrily — make the same trip hazard. The strong alternative to any tag is the action beat: a line of physical business anchored to the speaker that also advances the scene. "She set the cup down. 'You already knew.'" carries speaker, motion, and mood with no verb of speech at all.
Mechanics that hold up across edited fiction: place the tag early in a long speech so the reader knows whose mouth is open; use the em-dash interruption, not the ellipsis, when a character is cut off — the ellipsis means the speaker trailed off, which is a different event; and drop tags entirely once a two-person exchange has established its alternating rhythm, re-anchoring only when the rhythm breaks. These are conventions, not laws, but they solve the problems writers most often bring to workshops.
What is subtext, mechanically?
Characters talking about something other than the thing the scene is about. Ernest Hemingway's "Hills Like White Elephants" (1927) remains the classroom standard precisely because the entire argument — an unnamed operation the couple will not name — is conducted through a discussion of drinks, weather, and train times. The mechanics are visible: the characters' pressure shows in what they refuse to say, in subject changes that arrive one beat too fast, in agreement that is too quick.
To build it, write the scene's real conflict first in a paragraph of plain statement — she wants him to refuse; he wants her to choose — then give the characters a public subject to argue through, and let the real conflict bend every line without naming it. The paragraph is scaffolding and gets deleted. What remains is a scene with weather in it that is not about weather.
How do interruptions and overlap actually work on the page?
With restraint. The em-dash interruption—"I never said—" "You said it Tuesday."—reads as speed and pressure when it happens once or twice in a scene, and as noise when every exchange is cross-cut. Overlapping speech cannot truly be printed, so writers imply it: a character answers the question before it finishes, or starts a new topic while the old one is mid-sentence. Group scenes raise the difficulty, and the reliable technique is a traffic controller — one character whose lines steer who speaks next, keeping the reader's seat clear.
What does the repair bench look like?
| Symptom in the draft | Diagnosis | Fix |
|---|---|---|
| Lines all the same length | No moves, only exchange | Give each character a want; make one line change the other's position |
| Chortled, snapped, admonished | Tags doing emotion's work | Swap for "said" or an action beat carrying the feeling |
| Readers can't tell who's speaking | Shared nervous system | Build voice sheets; differentiate grammar habits, not vocabulary |
| Characters explain what both know | "As you know, brother" exposition | Move the information into conflict, or cut it and let context carry it |
| Every conversation agrees | No opposition | Find the line where the scene's want divides the speakers; end nearby |
| Accents phonetically spelled | Dialect written as spelling test | Suggest dialect through rhythm and word choice, not apostrophes |
How do you test dialogue by ear?
Read it aloud once per scene, but read it as the character, not as the narrator — actors' cold-reads catch what silent eyes forgive. Second test: the cover-the-names pass, already described, for voice separation. Third: the once-per-scene scan for the move — mark in the margin what each block of lines wanted and whether it got anything. A scene that survives all three tests still gets one more pass in revision, because dialogue is the part of prose where the ratio of problems found to pages examined runs highest. Writers who enjoy writing dialogue mostly enjoy having written it; the working method is the same enjoyment deferred.
Frequently asked questions
Should I ever use verbs other than "said"?
Occasionally, and mostly with purpose — "asked" after a question, or a pointed verb once a chapter where the surprise is the point. Leonard's rule is a default, not a statute; the failure is the ornament applied to every line, not the exception applied with intent.
How much real speech should I keep?
Only the fragments that carry character: one restart, one verbal tic, the unfinished sentence that shows the thought underneath. Greetings, weather, and small talk get cut unless the scene is secretly about them — the small talk doing subtext work stays precisely because it is a cover.
Does dialogue need dialect to feel real?
It needs rhythm more than spelling. Walter Scott Prize-level dialect transcription slows readers to a crawl; grammar and idiom differences — how a character forms questions, what contractions they avoid — do the work at full reading speed.
Can a whole scene be only dialogue, no tags?
Yes, once the alternation is established and the voices are distinct enough to survive the names-covered test. Re-anchor with an action beat every six to eight lines so the reader never loses their seat at the table.
For more context, read Why Opening Pages Get Rejected.
For more context, read strong verbs writing.
For more context, read Controlling Pacing With Sentence Length.