Three prompt iterations: how an AI fiction assistant fills in characters for the author

South African writer Coral Hart published 370 books in 20 months with AI assistance. That raises a question: what does a model actually do to fiction? The author of an analysis on Habr tested this with a scene from her own story: a man and a woman are playing chess, with a child nearby.

What happened on the first attempt

The prompt was short: it described the visible situation, stated that the man was confident of winning, asked for the characters' thoughts to be conveyed and specified a length of 1000–1500 characters. The BotHub assistant used the Claude Sonnet 5.5 preset, with a temperature of 0.8 and an output limit of 16 thousand tokens.

The model produced a solid scene in its own right. But it decided for itself that the child was the players' son, invented a shared past for the pair (“He used to wait for her to finish her book... Now he waited for her move”) and steered everything toward reconciliation. In the author's conception, the child is not their son at all, and the characters embody ambivalence. Put simply, the model filled the gaps with the most familiar plot.

What brought it closer to the author's intention

The author explained the hidden meaning to the model but asked it not to spell that meaning out for the reader. She then described the characters in greater detail: the man likes control but quickly understands the child; the woman is afraid of losing and tired of the game. She also had to remove the “self-contained vignette” effect: although the passage came from the middle of a story, the model kept reaching for a striking ending, so the prompt asked it to leave the game unfinished.

The third version introduced the intended contradiction: the man immediately notices a thread on the toy car's wheel but ignores the woman's wish to end the game. There were still flaws:

• The model interpreted “understands the child” literally, as being able to fix a toy, rather than deciphering the feelings behind a drawing;
• The line “You always leave me a choice that makes me guilty before I've even made it” explains the relationship too directly;
• There is a continuity error in the action: a hand hovers over the bishop, then repairs the toy car, and finally is “lifted off the piece,” even though the text never describes a return to the board.

The practical takeaway

This is not a benchmark, but one experiment by one author, without a blind reader test. Its takeaway is practical and applies to prose as well as to generative products more broadly.

• Spell out what the characters want and what stands in their way. A contradictory personality gives the model more to work with than the word “strict.”
• Supply what the reader does not yet know: the world's rules, relationships, the hidden intention and what should be revealed later.
• Specify the form: whose perspective to use, where the scene cuts off and what must remain unresolved.
• Check the continuity of actions and details; this remains a weakness even in successful versions.

The author also cites a 2024 study: stories written by people with suggestions from GPT-4 were rated as more original and enjoyable, but they were also more similar to one another. For product managers, this is an important signal: an AI assistant raises the average standard but narrows diversity. If you are building a tool for writers, let them specify hidden context and constraints, rather than just a topic and length. And make it clear that responsibility for the text remains with the person publishing it under their own name.