AI Dialogue Generator Tips: Getting Subtext, Distinct Voices and Silence Out of a Tool That Wants to Explain Everything

1. Name the real subject, then forbid it

The most effective single instruction. Every good scene of dialogue is about something the characters are not discussing. Tell the model what that is and tell it nobody may say it.

Situation: two sisters packing up their mother's kitchen.
Surface topic: which of them gets the good knives.
Real subject: which of them the mother loved more. Neither may say
this, name it, or allude to it directly. The reader should know it
by line ten.

Without the last two sentences, the model writes a sister saying "this was never about the knives, was it." With them, the knives carry the weight.

2. Give each speaker a voice sheet

Two characters in a generated scene sound alike because the prompt described them alike, or not at all. A minimal voice sheet per speaker:

JIN: short sentences, never asks a question, says "fine" when it is not,
     goes quiet rather than loud when angry.
MARA: long sentences that trail off, answers questions with questions,
     uses people's names when she is lying.

Then the test: strip the tags from the output and read it. If any line could be either sister, the sheets need one more rule each. The character prompts post has the full voice sheet prompt; the voice consistency guide covers keeping it stable across a book.

3. Ask for script format even for prose

A model writing prose dialogue narrates between the lines: "she said, her voice tight with the anger she'd been holding since the funeral." Ask for script format (name, colon, line; or screenplay cues) and that narration has nowhere to go. You get the lines alone, and you add the beats yourself, sparingly, in the conversion to prose. SparkLore's dialogue generator offers both output modes for this reason.

4. Forbid the explanatory line

Models resolve. Given tension, they write the line that releases it. A short banned list in the prompt:

Nobody may: state how they feel; explain why they did something;
summarise the past; tell the other person something they already know;
apologise; or agree. If a line does any of these, replace it with a
question, a change of subject, or an action.

The last clause matters: telling the model what to do instead of the banned move is what makes the ban work.

5. Specify the imbalance

Real conversations are unequal: one person wants something, one is withholding. Generated conversations are balanced, alternating lines of similar length. Fix it in the prompt:

Jin wants Mara to admit she took the money. Mara wants the
conversation to end. Jin speaks three lines for every one of Mara's.
Mara's lines get shorter as the scene goes on.

Line-count ratios are a crude instrument and they work.

6. Ask for silence and action as lines

A dialogue generator will fill every turn with words. Ask for the non-verbal turns:

At least three turns must be an action instead of a line (she puts a
knife in the box; he does not answer; she starts to say something and
stops). Mark them in parentheses. The scene ends on one of these.

The pause is the most powerful line in most scenes, and the model will never generate it unless it is asked to.

7. Use the halving pass

Generate, then run a second prompt on the output:

Cut this dialogue to half its lines. Delete lines that tell the other
person what they already know. Merge two lines with the same function.
Where a reply can be an action, delete the line and write the action.
Lose no fact and no relationship shift.

Almost every generated scene survives a 50% cut with nothing missing, and the survivors are the good lines. The prompt library has this as a form.

8. Break the scene in the wrong place

Models end dialogue scenes at the resolution. Ask for the end before it:

End the scene one line before either character gets what they want.

A scene that ends on the unanswered question carries the reader to the next one; the cliffhanger guide explains the mechanism at chapter scale, and it works at scene scale too.

9. Read for the tells

Generated dialogue has recognisable tics even when the above is applied. Before you keep a scene, look for:

• Name-calling in address. "Jin, I need you to understand." Real speakers rarely use the listener's name mid-conversation unless they are angry or lying; models use it every third line.
• Matched registers. Both speakers equally articulate. Real conversations have one person who is worse at this.
• The recap line. "After everything that happened last year..." The model summarising the plot for the reader.
• The therapeutic turn. A character naming their emotion in a full sentence. Nobody does this in a kitchen.
• The tidy exit. A last line that resolves or reconciles. Cut it.

The general AI-prose tells are in our editing guide; these are the dialogue-specific ones.

A worked prompt

Putting the tips together into a single prompt for a dialogue generator's free-text field:

Format: script (NAME: line). About twenty lines.
Situation: two sisters packing their mother's kitchen the day after
the funeral. Surface topic: the knives. Real subject: who was loved
more; nobody may say or allude to it.
JIN: short sentences, no questions, "fine" when not fine, quiet when angry.
MARA: trailing sentences, answers questions with questions, uses names
when lying. Jin speaks three lines to Mara's one; Mara's shorten.
Banned: stating feelings, explaining, recapping, apologising, agreeing.
Three turns are actions in parentheses. End one line before either
gets what she wants, on an action.

Then halve it, then convert to prose adding at most four beats. The scene that results is twelve lines long, is clearly about the mother, never mentions her, and could not be mistaken for a generated one, because the decisions in it were yours.

Try it in SparkLore

The free AI dialogue generator takes a situation and two characters and returns about twenty lines in prose or script mode, built to keep the real subject unspoken and the voices distinct; it runs three times a day without an account. The halving pass, the subtext argument and the "three voices" prompts are in the prompt library as forms. In the Studio, each character's voice sheet lives in the bible and is read whenever that character speaks. Sign up to keep the sheets with the manuscript; the pricing page has the plans.