How to Fix Generic Output With Better Taste, Better Inputs, and Better Systems

Refined editorial mood board with intentional materials, negative space, and curated visual references that contrast with generic generated output.

To fix generic output, stop blaming the model first and look at the workflow around it: a vague brief, no saved canon, taste nobody wrote down, and a habit of approving whatever looks most dramatic in the moment. Star Wars (1977) is a good case, because its look came from a written rule and from paintings that many of its shots copied.

Write your taste down

George Lucas had a rule for the film's look: "What is required for true credibility is a used future." Production designer Roger Christian said Lucas wanted everything to look "real and used." That is taste written as a sentence, and you can test a prop against it. The rebel ships look cobbled together next to the Empire's sleeker designs. The inside of the Millennium Falcon was built to look like a submarine, using cheap airplane scrap metal.

You can do the same for your book. Write down what the work should feel like, what would cheapen it, and which saved prompts and reference images keep giving you the right result. Say your story is a quiet, damp fishing town. Your page might say the mood is patient and salt-worn, that glossy postcard sunsets would cheapen it, and that three saved image prompts get the grey harbor right.

Give the model a real brief

A vague prompt gets the most average reading of the ask. That is why "a tense scene at the harbor" comes back looking like every other harbor scene. A useful brief says what the scene or image is for, how it should feel, and what must not appear.

Vader is the example here. Lucas wanted him to be "a very tall, dark fluttering figure." The script had Vader moving between ships, so Ralph McQuarrie gave him a kind of space suit, mixing a full-face breathing mask with a samurai helmet. The design answered what the scene needed. To avoid generic output in your own work, tell the tool where your villain has to walk, breathe, or fight, and let that shape the look.

References do heavy lifting too. In 1975 Lucas asked McQuarrie to paint several scenes from the script. Those paintings helped the bosses at 20th Century Fox picture the film. Lucas also made sure many shots copied the paintings exactly. Treat your approved images as that kind of picture, and feed them back in as reference. That is the core of better AI images, and it is real AI art direction rather than luck.

Keep your story together.

A manuscript studio with source-linked Wiki review.

Try AxiomStory free

Build a loop and check it

A creative AI workflow that works looks like this: settle canon, generate, check against your written rules, save what you approve with where it came from, and feed approved work back in as reference.

The check matters most. Approving the single most striking image or paragraph often backfires when it doesn't match the scenes next to it. Judge each new piece against what you have already approved.

We learned the checking part on our own tool. In November 2026 we checked the stories our story writing model wrote for signs of machine writing. Its habit wasn't the famous AI names; it kept reusing its own favourite. A character called "Mara" turned up in 5 of 6 unrelated stories, and about 42% of its paragraphs were a single short line. We added written rules to its prompts: a list of overused names to avoid, a hint for which letter each main character's name should start with, and style rules. Mara went from 5 stories to none, and one line paragraphs fell from 42% to 34%. A banned name, "Silas", still slipped through once, and our first letter hint made one story give all its main characters names starting with I. We changed the hint so each main character gets a different letter.

Our rough rule: any fix you write will make a new quirk, so read the results again after you add it.

AxiomStory makes pictures of your characters that keep each face the same from image to image, and gives each character a voice. It suggests edits you approve and never writes your book unless you ask. It works only in a browser and needs an internet connection.

Try this on your current project today. Pick five scenes you already like, write one sentence about what they share, then list what you would never allow. Hold your next generated page up against that sentence before you approve it.