Why
Subtext is not concealment for its own sake. It is the ordinary condition of people who want things from each other and have reasons not to say so.
The generative question is never "how do I make this less on the nose". It is what stops this character saying it plainly — and the answer produces the scene:
They do not know it. The audience sees it before they do. They know and will not admit it. Pride, shame, or love. They know and cannot say it here. Someone else is in the room. They know and saying it would lose them the thing they want.
Each reason produces a different kind of dialogue. The fourth is the most useful in a scene with an objective, because it ties the subtext directly to the want: speaking plainly would cost them the scene.
The failure in the other direction is worth naming. Dialogue with no surface — where nobody says anything about anything — is not subtext, it is evasion, and audiences find it as tiring as being told everything. Subtext requires a text: a real conversation, about a real topic, underneath which something else is happening.
The reliable structure is two conversations at once: one that both characters would describe if asked, and one neither would.
How
- Write the scene's subject — what the characters would say they are talking about. It should be concrete and mundane.
- Write what each actually wants. If either want is the same as the subject, the scene has no subtext yet.
- For each character, write the reason they cannot say the want plainly. If there is no reason, either add one or let them say it.
- Cut every line where a character names their own feeling. Replace with something they do to the other person.
- Check the audience can follow. Subtext the audience cannot decode is not subtext, and the usual fix is one small explicit signal early — after which everything can be indirect.
- Leave one line where somebody nearly says it and stops. That is the moment the scene is for.
Notes for AI film
Subtext is under-served by generated performance and over-served by generated framing. A model rendering a face will tend to render the emotion literally, which flattens exactly what this card is protecting. Carry the subtext in what is shown instead — hands, distance between bodies, who is looking away — and keep the face further from camera in the moment that matters most.
For text-driven formats, the caption is a legitimate second channel. What is written and what is shown can disagree, and that disagreement is subtext produced almost for free.