Ask AI only to “write a protagonist returning home after ten years,” and the result may invent rain, a dead mother, and an explanation of the protagonist’s mixed emotions. The passage may read smoothly, but none of those decisions are yours.
Before writing a longer prompt and asking for another complete passage, stop requesting prose. Ask AI to help locate what you have not decided.
Distinguish a description problem from a character-choice problem
You may think you are stuck describing the station when the real gap is why a protagonist who does not want to return steps off the train.
Write four short lines before opening AI:
- The scene you are writing
- The choice the character makes here
- What the reader must not know yet
- What you have not decided yourself
For example: “arrives at the home-town station / decides whether to leave the train / keep the truth about the fire hidden / reason for returning undecided.” This is enough to show what kind of help you need.
Ask for one question before asking for prose
Try an instruction like this:
Do not write the scene yet. I want to decide whether the protagonist leaves the train. Ask one question first about the character’s purpose or information still hidden from the reader. Wait for my answer before asking the next question, and do not infer an answer I have not given.
You do not need to answer every question. If you are not ready to decide why the protagonist originally left, leave it unresolved. A question that stops you may reveal that the obstacle was the character’s reason for acting, not the description of the station.
Define how alternatives should differ
“Give me three ideas” often produces three versions of the same idea. State the dimensions you want to compare: one where the protagonist chooses to confront the past, one where another person prompts the return, and one where an unexpected object changes the plan. Ask for the choice and the problem each alternative leaves for the next scene, not finished prose.
The most dramatic option is not automatically the right one. Explaining why you reject an option often exposes what the work needs to protect.
Test prose in small pieces beside your own draft
Once the direction is settled, request a passage small enough to compare—perhaps the entrance to a scene or one exchange of dialogue. Specify viewpoint, length, information that must remain hidden, and a sensory trigger. Compare the result with the existing manuscript:
- Does it reveal something the viewpoint character cannot know?
- Does the action follow from the character’s purpose?
- Does it silently change the timeline or setting?
- Does its vocabulary and rhythm belong beside the rest of the work?
A sentence can be attractive and still be wrong for the scene. That judgment does not need to be delegated.
Move adopted decisions out of chat
Long conversations keep accepted options, rejected options, and intermediate misunderstandings together. Put chosen prose in the manuscript. Put continuing facts—such as a character’s past or the fire’s chronology—in reference material. Keep unresolved questions explicitly unresolved.
Before sending an unpublished manuscript to any service, review how that service stores and uses input. Do not send someone else’s manuscript unless you have permission.
Progress does not have to mean counting how many words AI generated. If a previously hidden decision becomes clear and you can write the next paragraph yourself, the conversation has already done useful work.