Why I wouldn't generate incorrect quotes with the same prompt as a scene
Once a writing product has more than one format, it is tempting to reuse the scene prompt and add "keep it short." I would not do that for incorrect quotes. A quote is a different contract than a scene, and it needed its
Once a writing product has more than one format, it is tempting to reuse the scene prompt and add "keep it short." I would not do that for incorrect quotes. A quote is a different contract than a scene, and it needed its own inputs, its own output shape, and — as I wrote in a previous post — its own quota pool.
A quote is a different contract than a scene
A scene can ramble a little and still feel like a scene. An incorrect quote cannot. The format people actually share is short, labeled dialogue: two to four speakers, a few lines, and a punchline that changes how you read the lines before it. It should sound like the characters and also be something they would never say in canon.
If you ask a scene model to "make a short funny exchange," you usually get a mini-scene anyway: a setting sentence, a beat of action, then dialogue that over-explains the joke. The length instruction is treated as a hint. The model still tries to tell a story.
So I do not treat quotes as a shorter scene. The output contract is different: speaker labels, tight back-and-forth, no preamble, and a fixed number of quotes in one go (3, 5, or 10). A fluent mini-scene is still a failed quote.
The form is the prompt
A blank prompt box invites a paragraph of instructions, and then the model negotiates with itself. The quote tool does not work that way. The inputs are the constraints:
- A fandom, or an original world
- Two to four character names
- A relationship marker:
/for a ship,&for friends or family - Up to three tropes and a tone
- How many quotes to return
Those fields do more than decorate a UI. The relationship marker changes the joke. A slash sets up flirting that goes wrong; an ampersand sets up people who know each other too well. Tone does the rest. I would rather encode that in the form than hope a user types it into a prompt, and I would rather the model receive a structured request than a blob of prose. This is also why the character limit is small. Incorrect quotes fall apart when too many people are talking. Four names is already a crowd in three lines.
Don't guess pronouns
Short dialogue makes pronoun mistakes obvious. In a long scene, one wrong "she" can hide. In a four-line quote, it is the whole joke.
The generator takes pronouns in brackets after a name, for example Sam (they/them). Neopronouns and mixed pronouns are followed as typed: Ash (xe/xem), Rin (she/they). If a name has no pronouns, the tool does not infer gender from the name. It avoids gendered references to that character instead of guessing.
I would keep that rule even if it makes some lines a little more awkward. A wrong pronoun is worse than a slightly stiffer sentence, and guessing is a bug, not a feature.
It also needed its own quota pool
Writers often use a handful of quotes as a warm-up, or to test whether two voices bounce, before they spend a free scene. When every tool shared one daily counter, that warm-up could spend the only scene of the day. That is the failure I described in the quota post: a shared bucket is easy to explain and easy to regret.
Quotes and writing prompts are also cheaper to generate than a full scene, so they can have a higher cap once someone has bought credits — 50 sets a day on those short tools. A site-wide cap still sits on top, so a cheap surface cannot walk the model bill out the door.
The live tool is FanficWizard. Same product family as the scene generator; different contract.
What I would keep next time
If I added another format tomorrow — a prompt list, a character sheet, a one-line hook — I would give it its own output contract and its own pool on day one. "Make it shorter" is not a contract. A labeled, countable, pronoun-safe quote is.
I would also keep putting the important constraints in the form. Users should not have to learn my prompt dialect to get the right shape of answer.
Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.